merge with master

.gitattributes

@@ -0,0 +1,3 @@
+*.png filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -24,6 +24,7 @@ wheels/
 .installed.cfg
 *.egg
 /package
+/temp
 MANIFEST
 
 # PyInstaller
@@ -104,7 +105,6 @@ venv.bak/
 # mypy
 .mypy_cache/
 
-data
 .vscode
 .idea
 
@@ -124,3 +124,7 @@ replace.sh
 
 # Pytorch
 *.pth
+
+
+# audio
+*.wav

data/test/images/image1.jpg

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78094cc48fbcfd9b6d321fe13619ecc72b65e006fc1b4c4458409ade9979486d
+size 129862

data/test/images/image_captioning.png

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af83a94899a6d23339c3ecc5c4c58c57c835af57b531a2f4c50461184f820141
+size 603621

data/test/images/image_matting.png

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af83a94899a6d23339c3ecc5c4c58c57c835af57b531a2f4c50461184f820141
+size 603621

data/test/images/ocr_detection.jpg

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c8435db5583400be5d11a2c17910c96133b462c8a99ccaf0e19f4aac34e0a94
+size 141149

docs/source/develop.md

@@ -91,6 +91,55 @@ make tests
 
 4. Daily regression tests will run all cases at 0 am each day using master branch.
 
+### 2.3 Test data storage
+
+As we need a lot of data for testing, including images, videos, models. We use git lfs
+to store those large files.
+
+1. install git-lfs
+for mac
+```bash
+brew install git-lfs
+git lfs install
+```
+
+for centos, please download rpm from git-lfs github release [website](https://github.com/git-lfs/git-lfs/releases/tag/v3.2.0)
+```bash
+wget http://101374-public.oss-cn-hangzhou-zmf.aliyuncs.com/git-lfs-3.2.0-1.el7.x86_64.rpm
+sudo rpm -ivh git-lfs-3.2.0-1.el7.x86_64.rpm
+git lfs install
+```
+
+for ubuntu
+```bash
+curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
+sudo apt-get install git-lfs
+git lfs install
+```
+
+2. track your data type using git lfs, for example, to track png files
+```bash
+git lfs track "*.png"
+```
+
+3. add your test files to `data/test/` folder, you can make directories if you need.
+```bash
+git add data/test/test.png
+```
+
+4. commit your test data to remote branch
+```bash
+git commit -m "xxx"
+```
+
+To pull data from remote repo, just as the same way you pull git files.
+```bash
+git pull origin branch_name
+```
+
+
+
+
 ## Code Review
 
 1. Run following command to create an aone CR, replace `TARGET_BRANCH` and `CR_NAME` with the one you want.

docs/source/faq.md

@@ -29,3 +29,15 @@ reference: [https://huggingface.co/docs/tokenizers/installation#installation-fro
 > ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
 
 由于依赖库之间的版本不兼容，可能会存在版本冲突的情况，大部分情况下不影响正常运行。
+
+### 3. 安装pytorch出现版本错误
+
+> ERROR: Ignored the following versions that require a different python version: 1.1.0 Requires-Python >=3.8; 1.1.0rc1 Requires-Python >=3.8; 1.1.1 Requires-Python >=3.8
+> ERROR: Could not find a version that satisfies the requirement torch==1.8.1+cu111 (from versions: 1.0.0, 1.0.1, 1.0.1.post2, 1.1.0, 1.2.0, 1.3.0, 1.3.1, 1.4.0, 1.5.0, 1.5.1, 1.6.0, 1.7.0, 1.7.1, 1.8.0, 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.10.1, 1.10.2, 1.11.0)
+> ERROR: No matching distribution found for torch==1.8.1+cu111
+
+安装时使用如下命令：
+
+```shell
+pip install -f https://download.pytorch.org/whl/torch_stable.html -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt
+```

docs/source/quick_start.md

@@ -25,6 +25,10 @@ ModelScope Library目前支持tensorflow，pytorch两大深度学习框架进行
 * [Pytorch安装指导](https://pytorch.org/get-started/locally/)
 * [Tensorflow安装指导](https://www.tensorflow.org/install/pip)
 
+部分第三方依赖库需要提前安装numpy
+```
+pip install numpy
+```
 
 ## ModelScope library 安装
 

modelscope/models/__init__.py

@@ -1,5 +1,8 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
+from .audio.tts.am import SambertNetHifi16k
+from .audio.tts.vocoder import Hifigan16k
 from .base import Model
 from .builder import MODELS, build_model
+from .multi_model import OfaForImageCaptioning
 from .nlp import BertForSequenceClassification, SbertForSentenceSimilarity

modelscope/models/audio/layers/activations.py

@@ -0,0 +1,60 @@
+import torch.nn as nn
+
+from .layer_base import LayerBase
+
+
+class RectifiedLinear(LayerBase):
+
+    def __init__(self, input_dim, output_dim):
+        super(RectifiedLinear, self).__init__()
+        self.dim = input_dim
+        self.relu = nn.ReLU()
+
+    def forward(self, input):
+        return self.relu(input)
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<RectifiedLinear> %d %d\n' % (self.dim, self.dim)
+        return re_str
+
+    def load_kaldi_nnet(self, instr):
+        return instr
+
+
+class LogSoftmax(LayerBase):
+
+    def __init__(self, input_dim, output_dim):
+        super(LogSoftmax, self).__init__()
+        self.dim = input_dim
+        self.ls = nn.LogSoftmax()
+
+    def forward(self, input):
+        return self.ls(input)
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<Softmax> %d %d\n' % (self.dim, self.dim)
+        return re_str
+
+    def load_kaldi_nnet(self, instr):
+        return instr
+
+
+class Sigmoid(LayerBase):
+
+    def __init__(self, input_dim, output_dim):
+        super(Sigmoid, self).__init__()
+        self.dim = input_dim
+        self.sig = nn.Sigmoid()
+
+    def forward(self, input):
+        return self.sig(input)
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<Sigmoid> %d %d\n' % (self.dim, self.dim)
+        return re_str
+
+    def load_kaldi_nnet(self, instr):
+        return instr

modelscope/models/audio/layers/affine_transform.py

@@ -0,0 +1,78 @@
+import numpy as np
+import torch as th
+import torch.nn as nn
+
+from .layer_base import (LayerBase, expect_kaldi_matrix, expect_token_number,
+                         to_kaldi_matrix)
+
+
+class AffineTransform(LayerBase):
+
+    def __init__(self, input_dim, output_dim):
+        super(AffineTransform, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.linear = nn.Linear(input_dim, output_dim)
+
+    def forward(self, input):
+        return self.linear(input)
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<AffineTransform> %d %d\n' % (self.output_dim,
+                                                 self.input_dim)
+        re_str += '<LearnRateCoef> 1 <BiasLearnRateCoef> 1 <MaxNorm> 0\n'
+        linear_weights = self.state_dict()['linear.weight']
+        x = linear_weights.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        linear_bias = self.state_dict()['linear.bias']
+        x = linear_bias.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        return re_str
+
+    def to_raw_nnet(self, fid):
+        linear_weights = self.state_dict()['linear.weight']
+        x = linear_weights.squeeze().numpy()
+        x.tofile(fid)
+
+        linear_bias = self.state_dict()['linear.bias']
+        x = linear_bias.squeeze().numpy()
+        x.tofile(fid)
+
+    def load_kaldi_nnet(self, instr):
+        output = expect_token_number(
+            instr,
+            '<LearnRateCoef>',
+        )
+        if output is None:
+            raise Exception('AffineTransform format error for <LearnRateCoef>')
+        instr, lr = output
+
+        output = expect_token_number(instr, '<BiasLearnRateCoef>')
+        if output is None:
+            raise Exception(
+                'AffineTransform format error for <BiasLearnRateCoef>')
+        instr, lr = output
+
+        output = expect_token_number(instr, '<MaxNorm>')
+        if output is None:
+            raise Exception('AffineTransform format error for <MaxNorm>')
+        instr, lr = output
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('AffineTransform format error for parsing matrix')
+        instr, mat = output
+
+        print(mat.shape)
+        self.linear.weight = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('AffineTransform format error for parsing matrix')
+        instr, mat = output
+        mat = np.squeeze(mat)
+        self.linear.bias = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+        return instr

modelscope/models/audio/layers/deep_fsmn.py

@@ -0,0 +1,178 @@
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .layer_base import (LayerBase, expect_kaldi_matrix, expect_token_number,
+                         to_kaldi_matrix)
+
+
+class DeepFsmn(LayerBase):
+
+    def __init__(self,
+                 input_dim,
+                 output_dim,
+                 lorder=None,
+                 rorder=None,
+                 hidden_size=None,
+                 layer_norm=False,
+                 dropout=0):
+        super(DeepFsmn, self).__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+
+        if lorder is None:
+            return
+
+        self.lorder = lorder
+        self.rorder = rorder
+        self.hidden_size = hidden_size
+        self.layer_norm = layer_norm
+
+        self.linear = nn.Linear(input_dim, hidden_size)
+        self.norm = nn.LayerNorm(hidden_size)
+        self.drop1 = nn.Dropout(p=dropout)
+        self.drop2 = nn.Dropout(p=dropout)
+        self.project = nn.Linear(hidden_size, output_dim, bias=False)
+
+        self.conv1 = nn.Conv2d(
+            output_dim,
+            output_dim, [lorder, 1], [1, 1],
+            groups=output_dim,
+            bias=False)
+        self.conv2 = nn.Conv2d(
+            output_dim,
+            output_dim, [rorder, 1], [1, 1],
+            groups=output_dim,
+            bias=False)
+
+    def forward(self, input):
+
+        f1 = F.relu(self.linear(input))
+
+        f1 = self.drop1(f1)
+        if self.layer_norm:
+            f1 = self.norm(f1)
+
+        p1 = self.project(f1)
+
+        x = th.unsqueeze(p1, 1)
+
+        x_per = x.permute(0, 3, 2, 1)
+
+        y = F.pad(x_per, [0, 0, self.lorder - 1, 0])
+        yr = F.pad(x_per, [0, 0, 0, self.rorder])
+        yr = yr[:, :, 1:, :]
+
+        out = x_per + self.conv1(y) + self.conv2(yr)
+        out = self.drop2(out)
+
+        out1 = out.permute(0, 3, 2, 1)
+
+        return input + out1.squeeze()
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<UniDeepFsmn> %d %d\n'\
+                  % (self.output_dim, self.input_dim)
+        re_str += '<LearnRateCoef> %d <HidSize> %d <LOrder> %d <LStride> %d <MaxNorm> 0\n'\
+                  % (1, self.hidden_size, self.lorder, 1)
+        lfiters = self.state_dict()['conv1.weight']
+        x = np.flipud(lfiters.squeeze().numpy().T)
+        re_str += to_kaldi_matrix(x)
+        proj_weights = self.state_dict()['project.weight']
+        x = proj_weights.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        linear_weights = self.state_dict()['linear.weight']
+        x = linear_weights.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        linear_bias = self.state_dict()['linear.bias']
+        x = linear_bias.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        return re_str
+
+    def load_kaldi_nnet(self, instr):
+        output = expect_token_number(
+            instr,
+            '<LearnRateCoef>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <LearnRateCoef>')
+        instr, lr = output
+
+        output = expect_token_number(
+            instr,
+            '<HidSize>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <HidSize>')
+        instr, hiddensize = output
+        self.hidden_size = int(hiddensize)
+
+        output = expect_token_number(
+            instr,
+            '<LOrder>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <LOrder>')
+        instr, lorder = output
+        self.lorder = int(lorder)
+
+        output = expect_token_number(
+            instr,
+            '<LStride>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <LStride>')
+        instr, lstride = output
+        self.lstride = lstride
+
+        output = expect_token_number(
+            instr,
+            '<MaxNorm>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <MaxNorm>')
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+        mat1 = np.fliplr(mat.T).copy()
+        self.conv1 = nn.Conv2d(
+            self.output_dim,
+            self.output_dim, [self.lorder, 1], [1, 1],
+            groups=self.output_dim,
+            bias=False)
+        mat_th = th.from_numpy(mat1).type(th.FloatTensor)
+        mat_th = mat_th.unsqueeze(1)
+        mat_th = mat_th.unsqueeze(3)
+        self.conv1.weight = th.nn.Parameter(mat_th)
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+
+        self.project = nn.Linear(self.hidden_size, self.output_dim, bias=False)
+        self.linear = nn.Linear(self.input_dim, self.hidden_size)
+
+        self.project.weight = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+        self.linear.weight = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+        self.linear.bias = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        return instr

modelscope/models/audio/layers/layer_base.py

@@ -0,0 +1,50 @@
+import abc
+import re
+
+import numpy as np
+import torch.nn as nn
+
+
+def expect_token_number(instr, token):
+    first_token = re.match(r'^\s*' + token, instr)
+    if first_token is None:
+        return None
+    instr = instr[first_token.end():]
+    lr = re.match(r'^\s*(-?\d+\.?\d*e?-?\d*?)', instr)
+    if lr is None:
+        return None
+    return instr[lr.end():], lr.groups()[0]
+
+
+def expect_kaldi_matrix(instr):
+    pos2 = instr.find('[', 0)
+    pos3 = instr.find(']', pos2)
+    mat = []
+    for stt in instr[pos2 + 1:pos3].split('\n'):
+        tmp_mat = np.fromstring(stt, dtype=np.float32, sep=' ')
+        if tmp_mat.size > 0:
+            mat.append(tmp_mat)
+    return instr[pos3 + 1:], np.array(mat)
+
+
+def to_kaldi_matrix(np_mat):
+    """
+    function that transform as str numpy mat to standard kaldi str matrix
+    :param np_mat: numpy mat
+    :return: str
+    """
+    np.set_printoptions(threshold=np.inf, linewidth=np.nan, suppress=True)
+    out_str = str(np_mat)
+    out_str = out_str.replace('[', '')
+    out_str = out_str.replace(']', '')
+    return '[ %s ]\n' % out_str
+
+
+class LayerBase(nn.Module, metaclass=abc.ABCMeta):
+
+    def __init__(self):
+        super(LayerBase, self).__init__()
+
+    @abc.abstractmethod
+    def to_kaldi_nnet(self):
+        pass

modelscope/models/audio/layers/uni_deep_fsmn.py

@@ -0,0 +1,482 @@
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .layer_base import (LayerBase, expect_kaldi_matrix, expect_token_number,
+                         to_kaldi_matrix)
+
+
+class SepConv(nn.Module):
+
+    def __init__(self,
+                 in_channels,
+                 filters,
+                 out_channels,
+                 kernel_size=(5, 2),
+                 dilation=(1, 1)):
+        """ :param kernel_size (time, frequency)
+
+        """
+        super(SepConv, self).__init__()
+        # depthwise + pointwise
+        self.dconv = nn.Conv2d(
+            in_channels,
+            in_channels * filters,
+            kernel_size,
+            dilation=dilation,
+            groups=in_channels)
+        self.pconv = nn.Conv2d(
+            in_channels * filters, out_channels, kernel_size=1)
+        self.padding = dilation[0] * (kernel_size[0] - 1)
+
+    def forward(self, input):
+        ''' input: [B, C, T, F]
+        '''
+        x = F.pad(input, [0, 0, self.padding, 0])
+        x = self.dconv(x)
+        x = self.pconv(x)
+        return x
+
+
+class Conv2d(nn.Module):
+
+    def __init__(self,
+                 input_dim,
+                 output_dim,
+                 lorder=20,
+                 rorder=0,
+                 groups=1,
+                 bias=False,
+                 skip_connect=True):
+        super(Conv2d, self).__init__()
+        self.lorder = lorder
+        self.conv = nn.Conv2d(
+            input_dim, output_dim, [lorder, 1], groups=groups, bias=bias)
+        self.rorder = rorder
+        if self.rorder:
+            self.conv2 = nn.Conv2d(
+                input_dim, output_dim, [rorder, 1], groups=groups, bias=bias)
+        self.skip_connect = skip_connect
+
+    def forward(self, input):
+        # [B, 1, T, F]
+        x = th.unsqueeze(input, 1)
+        # [B, F, T, 1]
+        x_per = x.permute(0, 3, 2, 1)
+        y = F.pad(x_per, [0, 0, self.lorder - 1, 0])
+        out = self.conv(y)
+        if self.rorder:
+            yr = F.pad(x_per, [0, 0, 0, self.rorder])
+            yr = yr[:, :, 1:, :]
+            out += self.conv2(yr)
+        out = out.permute(0, 3, 2, 1).squeeze(1)
+        if self.skip_connect:
+            out = out + input
+        return out
+
+
+class SelfAttLayer(nn.Module):
+
+    def __init__(self, input_dim, output_dim, lorder=None, hidden_size=None):
+        super(SelfAttLayer, self).__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+
+        if lorder is None:
+            return
+
+        self.lorder = lorder
+        self.hidden_size = hidden_size
+
+        self.linear = nn.Linear(input_dim, hidden_size)
+
+        self.project = nn.Linear(hidden_size, output_dim, bias=False)
+
+        self.att = nn.Linear(input_dim, lorder, bias=False)
+
+    def forward(self, input):
+
+        f1 = F.relu(self.linear(input))
+
+        p1 = self.project(f1)
+
+        x = th.unsqueeze(p1, 1)
+
+        x_per = x.permute(0, 3, 2, 1)
+
+        y = F.pad(x_per, [0, 0, self.lorder - 1, 0])
+
+        # z [B, F, T, lorder]
+        z = x_per
+        for i in range(1, self.lorder):
+            z = th.cat([z, y[:, :, self.lorder - 1 - i:-i, :]], axis=-1)
+
+        # [B, T, lorder]
+        att = F.softmax(self.att(input), dim=-1)
+        att = th.unsqueeze(att, 1)
+        z = th.sum(z * att, axis=-1)
+
+        out1 = z.permute(0, 2, 1)
+
+        return input + out1
+
+
+class TFFsmn(nn.Module):
+
+    def __init__(self,
+                 input_dim,
+                 output_dim,
+                 lorder=None,
+                 hidden_size=None,
+                 dilation=1,
+                 layer_norm=False,
+                 dropout=0,
+                 skip_connect=True):
+        super(TFFsmn, self).__init__()
+
+        self.skip_connect = skip_connect
+
+        self.linear = nn.Linear(input_dim, hidden_size)
+        self.norm = nn.Identity()
+        if layer_norm:
+            self.norm = nn.LayerNorm(input_dim)
+        self.act = nn.ReLU()
+        self.project = nn.Linear(hidden_size, output_dim, bias=False)
+
+        self.conv1 = nn.Conv2d(
+            output_dim,
+            output_dim, [lorder, 1],
+            dilation=[dilation, 1],
+            groups=output_dim,
+            bias=False)
+        self.padding_left = dilation * (lorder - 1)
+        dorder = 5
+        self.conv2 = nn.Conv2d(1, 1, [dorder, 1], bias=False)
+        self.padding_freq = dorder - 1
+
+    def forward(self, input):
+        return self.compute1(input)
+
+    def compute1(self, input):
+        ''' linear-dconv-relu(norm)-linear-dconv
+        '''
+        x = self.linear(input)
+        # [B, 1, F, T]
+        x = th.unsqueeze(x, 1).permute(0, 1, 3, 2)
+        z = F.pad(x, [0, 0, self.padding_freq, 0])
+        z = self.conv2(z) + x
+        x = z.permute(0, 3, 2, 1).squeeze(-1)
+        x = self.act(x)
+        x = self.norm(x)
+        x = self.project(x)
+        x = th.unsqueeze(x, 1).permute(0, 3, 2, 1)
+        # [B, F, T+lorder-1, 1]
+        y = F.pad(x, [0, 0, self.padding_left, 0])
+        out = self.conv1(y)
+        if self.skip_connect:
+            out = out + x
+        out = out.permute(0, 3, 2, 1).squeeze()
+
+        return input + out
+
+
+class CNNFsmn(nn.Module):
+    ''' use cnn to reduce parameters
+    '''
+
+    def __init__(self,
+                 input_dim,
+                 output_dim,
+                 lorder=None,
+                 hidden_size=None,
+                 dilation=1,
+                 layer_norm=False,
+                 dropout=0,
+                 skip_connect=True):
+        super(CNNFsmn, self).__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.skip_connect = skip_connect
+
+        if lorder is None:
+            return
+
+        self.lorder = lorder
+        self.hidden_size = hidden_size
+
+        self.linear = nn.Linear(input_dim, hidden_size)
+        self.act = nn.ReLU()
+        kernel_size = (3, 8)
+        stride = (1, 4)
+        self.conv = nn.Sequential(
+            nn.ConstantPad2d((stride[1], 0, kernel_size[0] - 1, 0), 0),
+            nn.Conv2d(1, stride[1], kernel_size=kernel_size, stride=stride))
+
+        self.dconv = nn.Conv2d(
+            output_dim,
+            output_dim, [lorder, 1],
+            dilation=[dilation, 1],
+            groups=output_dim,
+            bias=False)
+        self.padding_left = dilation * (lorder - 1)
+
+    def forward(self, input):
+        return self.compute2(input)
+
+    def compute1(self, input):
+        ''' linear-relu(norm)-conv2d-relu?-dconv
+        '''
+        # [B, T, F]
+        x = self.linear(input)
+        x = self.act(x)
+        x = th.unsqueeze(x, 1)
+        x = self.conv(x)
+        # [B, C, T, F] -> [B, 1, T, F]
+        b, c, t, f = x.shape
+        x = x.view([b, 1, t, -1])
+        x = x.permute(0, 3, 2, 1)
+        # [B, F, T+lorder-1, 1]
+        y = F.pad(x, [0, 0, self.padding_left, 0])
+        out = self.dconv(y)
+        if self.skip_connect:
+            out = out + x
+        out = out.permute(0, 3, 2, 1).squeeze()
+        return input + out
+
+    def compute2(self, input):
+        ''' conv2d-relu-linear-relu?-dconv
+        '''
+        x = th.unsqueeze(input, 1)
+        x = self.conv(x)
+        x = self.act(x)
+        # [B, C, T, F] -> [B, T, F]
+        b, c, t, f = x.shape
+        x = x.view([b, t, -1])
+        x = self.linear(x)
+        x = th.unsqueeze(x, 1).permute(0, 3, 2, 1)
+        y = F.pad(x, [0, 0, self.padding_left, 0])
+        out = self.dconv(y)
+        if self.skip_connect:
+            out = out + x
+        out = out.permute(0, 3, 2, 1).squeeze()
+        return input + out
+
+
+class UniDeepFsmn(LayerBase):
+
+    def __init__(self,
+                 input_dim,
+                 output_dim,
+                 lorder=None,
+                 hidden_size=None,
+                 dilation=1,
+                 layer_norm=False,
+                 dropout=0,
+                 skip_connect=True):
+        super(UniDeepFsmn, self).__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.skip_connect = skip_connect
+
+        if lorder is None:
+            return
+
+        self.lorder = lorder
+        self.hidden_size = hidden_size
+
+        self.linear = nn.Linear(input_dim, hidden_size)
+        self.norm = nn.Identity()
+        if layer_norm:
+            self.norm = nn.LayerNorm(input_dim)
+        self.act = nn.ReLU()
+        self.project = nn.Linear(hidden_size, output_dim, bias=False)
+
+        self.conv1 = nn.Conv2d(
+            output_dim,
+            output_dim, [lorder, 1],
+            dilation=[dilation, 1],
+            groups=output_dim,
+            bias=False)
+        self.padding_left = dilation * (lorder - 1)
+
+    def forward(self, input):
+        return self.compute1(input)
+
+    def compute1(self, input):
+        ''' linear-relu(norm)-linear-dconv
+        '''
+        # [B, T, F]
+        x = self.linear(input)
+        x = self.act(x)
+        x = self.norm(x)
+        x = self.project(x)
+        x = th.unsqueeze(x, 1).permute(0, 3, 2, 1)
+        # [B, F, T+lorder-1, 1]
+        y = F.pad(x, [0, 0, self.padding_left, 0])
+        out = self.conv1(y)
+        if self.skip_connect:
+            out = out + x
+        out = out.permute(0, 3, 2, 1).squeeze()
+
+        return input + out
+
+    def compute2(self, input):
+        ''' linear-dconv-linear-relu(norm)
+        '''
+        x = self.project(input)
+        x = th.unsqueeze(x, 1).permute(0, 3, 2, 1)
+        y = F.pad(x, [0, 0, self.padding_left, 0])
+        out = self.conv1(y)
+        if self.skip_connect:
+            out = out + x
+        out = out.permute(0, 3, 2, 1).squeeze()
+        x = self.linear(out)
+        x = self.act(x)
+        x = self.norm(x)
+
+        return input + x
+
+    def compute3(self, input):
+        ''' dconv-linear-relu(norm)-linear
+        '''
+        x = th.unsqueeze(input, 1).permute(0, 3, 2, 1)
+        y = F.pad(x, [0, 0, self.padding_left, 0])
+        out = self.conv1(y)
+        if self.skip_connect:
+            out = out + x
+        out = out.permute(0, 3, 2, 1).squeeze()
+        x = self.linear(out)
+        x = self.act(x)
+        x = self.norm(x)
+        x = self.project(x)
+
+        return input + x
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<UniDeepFsmn> %d %d\n' \
+                  % (self.output_dim, self.input_dim)
+        re_str += '<LearnRateCoef> %d <HidSize> %d <LOrder> %d <LStride> %d <MaxNorm> 0\n' \
+                  % (1, self.hidden_size, self.lorder, 1)
+        lfiters = self.state_dict()['conv1.weight']
+        x = np.flipud(lfiters.squeeze().numpy().T)
+        re_str += to_kaldi_matrix(x)
+        proj_weights = self.state_dict()['project.weight']
+        x = proj_weights.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        linear_weights = self.state_dict()['linear.weight']
+        x = linear_weights.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        linear_bias = self.state_dict()['linear.bias']
+        x = linear_bias.squeeze().numpy()
+        re_str += to_kaldi_matrix(x)
+        return re_str
+
+    def to_raw_nnet(self, fid):
+        lfiters = self.state_dict()['conv1.weight']
+        x = np.flipud(lfiters.squeeze().numpy().T)
+        x.tofile(fid)
+
+        proj_weights = self.state_dict()['project.weight']
+        x = proj_weights.squeeze().numpy()
+        x.tofile(fid)
+
+        linear_weights = self.state_dict()['linear.weight']
+        x = linear_weights.squeeze().numpy()
+        x.tofile(fid)
+
+        linear_bias = self.state_dict()['linear.bias']
+        x = linear_bias.squeeze().numpy()
+        x.tofile(fid)
+
+    def load_kaldi_nnet(self, instr):
+        output = expect_token_number(
+            instr,
+            '<LearnRateCoef>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <LearnRateCoef>')
+        instr, lr = output
+
+        output = expect_token_number(
+            instr,
+            '<HidSize>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <HidSize>')
+        instr, hiddensize = output
+        self.hidden_size = int(hiddensize)
+
+        output = expect_token_number(
+            instr,
+            '<LOrder>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <LOrder>')
+        instr, lorder = output
+        self.lorder = int(lorder)
+
+        output = expect_token_number(
+            instr,
+            '<LStride>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <LStride>')
+        instr, lstride = output
+        self.lstride = lstride
+
+        output = expect_token_number(
+            instr,
+            '<MaxNorm>',
+        )
+        if output is None:
+            raise Exception('UniDeepFsmn format error for <MaxNorm>')
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+        mat1 = np.fliplr(mat.T).copy()
+
+        self.conv1 = nn.Conv2d(
+            self.output_dim,
+            self.output_dim, [self.lorder, 1], [1, 1],
+            groups=self.output_dim,
+            bias=False)
+
+        mat_th = th.from_numpy(mat1).type(th.FloatTensor)
+        mat_th = mat_th.unsqueeze(1)
+        mat_th = mat_th.unsqueeze(3)
+        self.conv1.weight = th.nn.Parameter(mat_th)
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+
+        self.project = nn.Linear(self.hidden_size, self.output_dim, bias=False)
+        self.linear = nn.Linear(self.input_dim, self.hidden_size)
+
+        self.project.weight = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+        self.linear.weight = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        output = expect_kaldi_matrix(instr)
+        if output is None:
+            raise Exception('UniDeepFsmn format error for parsing matrix')
+        instr, mat = output
+        mat = np.squeeze(mat)
+        self.linear.bias = th.nn.Parameter(
+            th.from_numpy(mat).type(th.FloatTensor))
+
+        return instr

modelscope/models/audio/network/loss.py

@@ -0,0 +1,394 @@
+import torch
+import torch.nn.functional as F
+
+from .modulation_loss import (GaborSTRFConv, MelScale,
+                              ModulationDomainLossModule)
+
+EPS = 1e-8
+
+
+def compute_mask(mixed_spec, clean_spec, mask_type='psmiam', clip=1):
+    '''
+        stft: (batch, ..., 2) or complex(batch, ...)
+        y = x + n
+    '''
+    if torch.is_complex(mixed_spec):
+        yr, yi = mixed_spec.real, mixed_spec.imag
+    else:
+        yr, yi = mixed_spec[..., 0], mixed_spec[..., 1]
+    if torch.is_complex(clean_spec):
+        xr, xi = clean_spec.real, clean_spec.imag
+    else:
+        xr, xi = clean_spec[..., 0], clean_spec[..., 1]
+
+    if mask_type == 'iam':
+        ymag = torch.sqrt(yr**2 + yi**2)
+        xmag = torch.sqrt(xr**2 + xi**2)
+        iam = xmag / (ymag + EPS)
+        return torch.clamp(iam, 0, 1)
+
+    elif mask_type == 'psm':
+        ypow = yr**2 + yi**2
+        psm = (xr * yr + xi * yi) / (ypow + EPS)
+        return torch.clamp(psm, 0, 1)
+
+    elif mask_type == 'psmiam':
+        ypow = yr**2 + yi**2
+        psm = (xr * yr + xi * yi) / (ypow + EPS)
+        ymag = torch.sqrt(yr**2 + yi**2)
+        xmag = torch.sqrt(xr**2 + xi**2)
+        iam = xmag / (ymag + EPS)
+        psmiam = psm * iam
+        return torch.clamp(psmiam, 0, 1)
+
+    elif mask_type == 'crm':
+        ypow = yr**2 + yi**2
+        mr = (xr * yr + xi * yi) / (ypow + EPS)
+        mi = (xi * yr - xr * yi) / (ypow + EPS)
+        mr = torch.clamp(mr, -clip, clip)
+        mi = torch.clamp(mi, -clip, clip)
+        return mr, mi
+
+
+def energy_vad(spec,
+               thdhigh=320 * 600 * 600 * 2,
+               thdlow=320 * 300 * 300 * 2,
+               int16=True):
+    '''
+        energy based vad should be accurate enough
+        spec: (batch, bins, frames, 2)
+        returns (batch, frames)
+    '''
+    energy = torch.sum(spec[..., 0]**2 + spec[..., 1]**2, dim=1)
+    vad = energy > thdhigh
+    idx = torch.logical_and(vad == 0, energy > thdlow)
+    vad[idx] = 0.5
+    return vad
+
+
+def modulation_loss_init(n_fft):
+    gabor_strf_parameters = torch.load(
+        './network/gabor_strf_parameters.pt')['state_dict']
+    gabor_modulation_kernels = GaborSTRFConv(supn=30, supk=30, nkern=60)
+    gabor_modulation_kernels.load_state_dict(gabor_strf_parameters)
+
+    modulation_loss_module = ModulationDomainLossModule(
+        gabor_modulation_kernels.eval())
+    for param in modulation_loss_module.parameters():
+        param.requires_grad = False
+
+    stft2mel = MelScale(
+        n_mels=80, sample_rate=16000, n_stft=n_fft // 2 + 1).cuda()
+
+    return modulation_loss_module, stft2mel
+
+
+def mask_loss_function(
+        loss_func='psm_loss',
+        loss_type='mse',  # ['mse', 'mae', 'comb']
+        mask_type='psmiam',
+        use_mod_loss=False,
+        use_wav2vec_loss=False,
+        n_fft=640,
+        hop_length=320,
+        EPS=1e-8,
+        weight=None):
+    if weight is not None:
+        print(f'Use loss weight: {weight}')
+    winlen = n_fft
+    window = torch.hamming_window(winlen, periodic=False)
+
+    def stft(x, return_complex=False):
+        # returns [batch, bins, frames, 2]
+        return torch.stft(
+            x,
+            n_fft,
+            hop_length,
+            winlen,
+            window=window.to(x.device),
+            center=False,
+            return_complex=return_complex)
+
+    def istft(x, slen):
+        return torch.istft(
+            x,
+            n_fft,
+            hop_length,
+            winlen,
+            window=window.to(x.device),
+            center=False,
+            length=slen)
+
+    def mask_loss(targets, masks, nframes):
+        ''' [Batch, Time, Frequency]
+        '''
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(targets)
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        masks = masks * mask_for_loss
+        targets = targets * mask_for_loss
+
+        if weight is None:
+            alpha = 1
+        else:  # for aec ST
+            alpha = weight - targets
+
+        if loss_type == 'mse':
+            loss = 0.5 * torch.sum(alpha * torch.pow(targets - masks, 2))
+        elif loss_type == 'mae':
+            loss = torch.sum(alpha * torch.abs(targets - masks))
+        else:  # mse(mask), mae(mask) approx 1:2
+            loss = 0.5 * torch.sum(alpha * torch.pow(targets - masks, 2)
+                                   + 0.1 * alpha * torch.abs(targets - masks))
+        loss /= torch.sum(nframes)
+        return loss
+
+    def spectrum_loss(targets, spec, nframes):
+        ''' [Batch, Time, Frequency, 2]
+        '''
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(targets[..., 0])
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        xr = spec[..., 0] * mask_for_loss
+        xi = spec[..., 1] * mask_for_loss
+        yr = targets[..., 0] * mask_for_loss
+        yi = targets[..., 1] * mask_for_loss
+        xmag = torch.sqrt(spec[..., 0]**2 + spec[..., 1]**2) * mask_for_loss
+        ymag = torch.sqrt(targets[..., 0]**2
+                          + targets[..., 1]**2) * mask_for_loss
+
+        loss1 = torch.sum(torch.pow(xr - yr, 2) + torch.pow(xi - yi, 2))
+        loss2 = torch.sum(torch.pow(xmag - ymag, 2))
+
+        loss = (loss1 + loss2) / torch.sum(nframes)
+        return loss
+
+    def sa_loss_dlen(mixed, clean, masks, nframes):
+        yspec = stft(mixed).permute([0, 2, 1, 3]) / 32768
+        xspec = stft(clean).permute([0, 2, 1, 3]) / 32768
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(xspec[..., 0])
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        emag = ((yspec[..., 0]**2 + yspec[..., 1]**2)**0.15) * (masks**0.3)
+        xmag = (xspec[..., 0]**2 + xspec[..., 1]**2)**0.15
+        emag = emag * mask_for_loss
+        xmag = xmag * mask_for_loss
+
+        loss = torch.sum(torch.pow(emag - xmag, 2)) / torch.sum(nframes)
+        return loss
+
+    def psm_vad_loss_dlen(mixed, clean, masks, nframes, subtask=None):
+        mixed_spec = stft(mixed)
+        clean_spec = stft(clean)
+        targets = compute_mask(mixed_spec, clean_spec, mask_type)
+        # [B, T, F]
+        targets = targets.permute(0, 2, 1)
+
+        loss = mask_loss(targets, masks, nframes)
+
+        if subtask is not None:
+            vadtargets = energy_vad(clean_spec)
+            with torch.no_grad():
+                mask_for_loss = torch.ones_like(targets[:, :, 0])
+                for idx, num in enumerate(nframes):
+                    mask_for_loss[idx, num:] = 0
+            subtask = subtask[:, :, 0] * mask_for_loss
+            vadtargets = vadtargets * mask_for_loss
+
+            loss_vad = F.binary_cross_entropy(subtask, vadtargets)
+            return loss + loss_vad
+        return loss
+
+    def modulation_loss(mixed, clean, masks, nframes, subtask=None):
+        mixed_spec = stft(mixed, True)
+        clean_spec = stft(clean, True)
+        enhanced_mag = torch.abs(mixed_spec)
+        clean_mag = torch.abs(clean_spec)
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(clean_mag)
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, :, num:] = 0
+        clean_mag = clean_mag * mask_for_loss
+        enhanced_mag = enhanced_mag * mask_for_loss * masks.permute([0, 2, 1])
+
+        # Covert to log-mel representation
+        # (B,T,#mel_channels)
+        clean_log_mel = torch.log(
+            torch.transpose(stft2mel(clean_mag**2), 2, 1) + 1e-8)
+        enhanced_log_mel = torch.log(
+            torch.transpose(stft2mel(enhanced_mag**2), 2, 1) + 1e-8)
+
+        alpha = compute_mask(mixed_spec, clean_spec, mask_type)
+        alpha = alpha.permute(0, 2, 1)
+        loss = 0.05 * modulation_loss_module(enhanced_log_mel, clean_log_mel,
+                                             alpha)
+        loss2 = psm_vad_loss_dlen(mixed, clean, masks, nframes, subtask)
+        # print(loss.item(), loss2.item()) #approx 1:4
+        loss = loss + loss2
+        return loss
+
+    def wav2vec_loss(mixed, clean, masks, nframes, subtask=None):
+        mixed /= 32768
+        clean /= 32768
+        mixed_spec = stft(mixed)
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(masks)
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        masks_est = masks * mask_for_loss
+
+        estimate = mixed_spec * masks_est.permute([0, 2, 1]).unsqueeze(3)
+        est_clean = istft(estimate, clean.shape[1])
+        loss = wav2vec_loss_module(est_clean, clean)
+        return loss
+
+    def sisdr_loss_dlen(mixed,
+                        clean,
+                        masks,
+                        nframes,
+                        subtask=None,
+                        zero_mean=True):
+        mixed_spec = stft(mixed)
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(masks)
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        masks_est = masks * mask_for_loss
+
+        estimate = mixed_spec * masks_est.permute([0, 2, 1]).unsqueeze(3)
+        est_clean = istft(estimate, clean.shape[1])
+        flen = min(clean.shape[1], est_clean.shape[1])
+        clean = clean[:, :flen]
+        est_clean = est_clean[:, :flen]
+
+        # follow asteroid/losses/sdr.py
+        if zero_mean:
+            clean = clean - torch.mean(clean, dim=1, keepdim=True)
+            est_clean = est_clean - torch.mean(est_clean, dim=1, keepdim=True)
+
+        dot = torch.sum(est_clean * clean, dim=1, keepdim=True)
+        s_clean_energy = torch.sum(clean**2, dim=1, keepdim=True) + EPS
+        scaled_clean = dot * clean / s_clean_energy
+        e_noise = est_clean - scaled_clean
+
+        # [batch]
+        sisdr = torch.sum(
+            scaled_clean**2, dim=1) / (
+                torch.sum(e_noise**2, dim=1) + EPS)
+        sisdr = -10 * torch.log10(sisdr + EPS)
+        loss = sisdr.mean()
+        return loss
+
+    def sisdr_freq_loss_dlen(mixed, clean, masks, nframes, subtask=None):
+        mixed_spec = stft(mixed)
+        clean_spec = stft(clean)
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(masks)
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        masks_est = masks * mask_for_loss
+
+        estimate = mixed_spec * masks_est.permute([0, 2, 1]).unsqueeze(3)
+
+        dot_real = estimate[..., 0] * clean_spec[..., 0] + \
+            estimate[..., 1] * clean_spec[..., 1]
+        dot_imag = estimate[..., 0] * clean_spec[..., 1] - \
+            estimate[..., 1] * clean_spec[..., 0]
+        dot = torch.cat([dot_real.unsqueeze(3), dot_imag.unsqueeze(3)], dim=-1)
+        s_clean_energy = clean_spec[..., 0] ** 2 + \
+            clean_spec[..., 1] ** 2 + EPS
+        scaled_clean = dot * clean_spec / s_clean_energy.unsqueeze(3)
+        e_noise = estimate - scaled_clean
+
+        # [batch]
+        scaled_clean_energy = torch.sum(
+            scaled_clean[..., 0]**2 + scaled_clean[..., 1]**2, dim=1)
+        e_noise_energy = torch.sum(
+            e_noise[..., 0]**2 + e_noise[..., 1]**2, dim=1)
+        sisdr = torch.sum(
+            scaled_clean_energy, dim=1) / (
+                torch.sum(e_noise_energy, dim=1) + EPS)
+        sisdr = -10 * torch.log10(sisdr + EPS)
+        loss = sisdr.mean()
+        return loss
+
+    def crm_loss_dlen(mixed, clean, masks, nframes, subtask=None):
+        mixed_spec = stft(mixed).permute([0, 2, 1, 3])
+        clean_spec = stft(clean).permute([0, 2, 1, 3])
+        mixed_spec = mixed_spec / 32768
+        clean_spec = clean_spec / 32768
+        tgt_mr, tgt_mi = compute_mask(mixed_spec, clean_spec, mask_type='crm')
+
+        D = int(masks.shape[2] / 2)
+        with torch.no_grad():
+            mask_for_loss = torch.ones_like(clean_spec[..., 0])
+            for idx, num in enumerate(nframes):
+                mask_for_loss[idx, num:, :] = 0
+        mr = masks[..., :D] * mask_for_loss
+        mi = masks[..., D:] * mask_for_loss
+        tgt_mr = tgt_mr * mask_for_loss
+        tgt_mi = tgt_mi * mask_for_loss
+
+        if weight is None:
+            alpha = 1
+        else:
+            alpha = weight - tgt_mr
+        # signal approximation
+        yr = mixed_spec[..., 0]
+        yi = mixed_spec[..., 1]
+        loss1 = torch.sum(alpha * torch.pow((mr * yr - mi * yi) - clean_spec[..., 0], 2)) \
+            + torch.sum(alpha * torch.pow((mr * yi + mi * yr) - clean_spec[..., 1], 2))
+        # mask approximation
+        loss2 = torch.sum(alpha * torch.pow(mr - tgt_mr, 2)) \
+            + torch.sum(alpha * torch.pow(mi - tgt_mi, 2))
+        loss = 0.5 * (loss1 + loss2) / torch.sum(nframes)
+        return loss
+
+    def crm_miso_loss_dlen(mixed, clean, masks, nframes):
+        return crm_loss_dlen(mixed[..., 0], clean[..., 0], masks, nframes)
+
+    def mimo_loss_dlen(mixed, clean, masks, nframes):
+        chs = mixed.shape[-1]
+        D = masks.shape[2] // chs
+        loss = psm_vad_loss_dlen(mixed[..., 0], clean[..., 0], masks[..., :D],
+                                 nframes)
+        for ch in range(1, chs):
+            loss1 = psm_vad_loss_dlen(mixed[..., ch], clean[..., ch],
+                                      masks[..., ch * D:ch * D + D], nframes)
+            loss = loss + loss1
+        return loss / chs
+
+    def spec_loss_dlen(mixed, clean, spec, nframes):
+        clean_spec = stft(clean).permute([0, 2, 1, 3])
+        clean_spec = clean_spec / 32768
+
+        D = spec.shape[2] // 2
+        spec_est = torch.cat([spec[..., :D, None], spec[..., D:, None]],
+                             dim=-1)
+        loss = spectrum_loss(clean_spec, spec_est, nframes)
+        return loss
+
+    if loss_func == 'psm_vad_loss_dlen':
+        return psm_vad_loss_dlen
+    elif loss_func == 'sisdr_loss_dlen':
+        return sisdr_loss_dlen
+    elif loss_func == 'sisdr_freq_loss_dlen':
+        return sisdr_freq_loss_dlen
+    elif loss_func == 'crm_loss_dlen':
+        return crm_loss_dlen
+    elif loss_func == 'modulation_loss':
+        return modulation_loss
+    elif loss_func == 'wav2vec_loss':
+        return wav2vec_loss
+    elif loss_func == 'mimo_loss_dlen':
+        return mimo_loss_dlen
+    elif loss_func == 'spec_loss_dlen':
+        return spec_loss_dlen
+    elif loss_func == 'sa_loss_dlen':
+        return sa_loss_dlen
+    else:
+        print('error loss func')
+        return None

modelscope/models/audio/network/modulation_loss.py

@@ -0,0 +1,248 @@
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchaudio.transforms import MelScale
+
+
+class ModulationDomainLossModule(torch.nn.Module):
+    """Modulation-domain loss function developed in [1] for supervised speech enhancement
+
+        In our paper, we used the gabor-based STRF kernels as the modulation kernels and used the log-mel spectrogram
+        as the input spectrogram representation.
+        Specific parameter details are in the paper and in the example below
+
+        Parameters
+        ----------
+        modulation_kernels: nn.Module
+            Differentiable module that transforms a spectrogram representation to the modulation domain
+
+            modulation_domain = modulation_kernels(input_tf_representation)
+            Input Spectrogram representation (B, T, F) ---> |(M) modulation_kernels|--->Modulation Domain(B, M, T', F')
+
+        norm: boolean
+            Normalizes the modulation domain representation to be 0 mean across time
+
+        [1] T. Vuong, Y. Xia, and R. M. Stern, “A modulation-domain lossfor neural-network-based real-time
+         speech enhancement”
+            Accepted ICASSP 2021, https://arxiv.org/abs/2102.07330
+
+
+    """
+
+    def __init__(self, modulation_kernels, norm=True):
+        super(ModulationDomainLossModule, self).__init__()
+
+        self.modulation_kernels = modulation_kernels
+        self.mse = nn.MSELoss(reduce=False)
+        self.norm = norm
+
+    def forward(self, enhanced_spect, clean_spect, weight=None):
+        """Calculate modulation-domain loss
+        Args:
+            enhanced_spect (Tensor): spectrogram representation of enhanced signal (B, #frames, #freq_channels).
+            clean_spect (Tensor): spectrogram representation of clean ground-truth signal (B, #frames, #freq_channels).
+        Returns:
+            Tensor: Modulation-domain loss value.
+        """
+
+        clean_mod = self.modulation_kernels(clean_spect)
+        enhanced_mod = self.modulation_kernels(enhanced_spect)
+
+        if self.norm:
+            mean_clean_mod = torch.mean(clean_mod, dim=2)
+            mean_enhanced_mod = torch.mean(enhanced_mod, dim=2)
+
+            clean_mod = clean_mod - mean_clean_mod.unsqueeze(2)
+            enhanced_mod = enhanced_mod - mean_enhanced_mod.unsqueeze(2)
+
+        if weight is None:
+            alpha = 1
+        else:  # TF-mask weight
+            alpha = 1 + torch.sum(weight, dim=-1, keepdim=True).unsqueeze(1)
+        mod_mse_loss = self.mse(enhanced_mod, clean_mod) * alpha
+        mod_mse_loss = torch.mean(
+            torch.sum(mod_mse_loss, dim=(1, 2, 3))
+            / torch.sum(clean_mod**2, dim=(1, 2, 3)))
+
+        return mod_mse_loss
+
+
+class ModulationDomainNCCLossModule(torch.nn.Module):
+    """Modulation-domain loss function developed in [1] for supervised speech enhancement
+
+        # Speech Intelligibility Prediction Using Spectro-Temporal Modulation Analysis - based off of this
+
+        In our paper, we used the gabor-based STRF kernels as the modulation kernels and used the log-mel spectrogram
+        as the input spectrogram representation.
+        Specific parameter details are in the paper and in the example below
+
+        Parameters
+        ----------
+        modulation_kernels: nn.Module
+            Differentiable module that transforms a spectrogram representation to the modulation domain
+
+            modulation_domain = modulation_kernels(input_tf_representation)
+            Input Spectrogram representation(B, T, F) --- (M) modulation_kernels---> Modulation Domain(B, M, T', F')
+
+        [1]
+
+    """
+
+    def __init__(self, modulation_kernels):
+        super(ModulationDomainNCCLossModule, self).__init__()
+
+        self.modulation_kernels = modulation_kernels
+        self.mse = nn.MSELoss(reduce=False)
+
+    def forward(self, enhanced_spect, clean_spect):
+        """Calculate modulation-domain loss
+        Args:
+            enhanced_spect (Tensor): spectrogram representation of enhanced signal (B, #frames, #freq_channels).
+            clean_spect (Tensor): spectrogram representation of clean ground-truth signal (B, #frames, #freq_channels).
+        Returns:
+            Tensor: Modulation-domain loss value.
+        """
+
+        clean_mod = self.modulation_kernels(clean_spect)
+        enhanced_mod = self.modulation_kernels(enhanced_spect)
+        mean_clean_mod = torch.mean(clean_mod, dim=2)
+        mean_enhanced_mod = torch.mean(enhanced_mod, dim=2)
+
+        normalized_clean = clean_mod - mean_clean_mod.unsqueeze(2)
+        normalized_enhanced = enhanced_mod - mean_enhanced_mod.unsqueeze(2)
+
+        inner_product = torch.sum(
+            normalized_clean * normalized_enhanced, dim=2)
+        normalized_denom = (torch.sum(
+            normalized_clean * normalized_clean, dim=2))**.5 * (torch.sum(
+                normalized_enhanced * normalized_enhanced, dim=2))**.5
+
+        ncc = inner_product / normalized_denom
+        mod_mse_loss = torch.mean((ncc - 1.0)**2)
+
+        return mod_mse_loss
+
+
+class GaborSTRFConv(nn.Module):
+    """Gabor-STRF-based cross-correlation kernel."""
+
+    def __init__(self,
+                 supn,
+                 supk,
+                 nkern,
+                 rates=None,
+                 scales=None,
+                 norm_strf=True,
+                 real_only=False):
+        """Instantiate a Gabor-based STRF convolution layer.
+        Parameters
+        ----------
+        supn: int
+            Time support in number of frames. Also the window length.
+        supk: int
+            Frequency support in number of channels. Also the window length.
+        nkern: int
+            Number of kernels, each with a learnable rate and scale.
+        rates: list of float, None
+            Initial values for temporal modulation.
+        scales: list of float, None
+            Initial values for spectral modulation.
+        norm_strf: Boolean
+            Normalize STRF kernels to be unit length
+        real_only: Boolean
+            If True, nkern REAL gabor-STRF kernels
+            If False, nkern//2 REAL and nkern//2 IMAGINARY gabor-STRF kernels
+        """
+        super(GaborSTRFConv, self).__init__()
+        self.numN = supn
+        self.numK = supk
+        self.numKern = nkern
+        self.real_only = real_only
+        self.norm_strf = norm_strf
+
+        if not real_only:
+            nkern = nkern // 2
+
+        if supk % 2 == 0:  # force odd number
+            supk += 1
+        self.supk = torch.arange(supk, dtype=torch.float32)
+        if supn % 2 == 0:  # force odd number
+            supn += 1
+        self.supn = torch.arange(supn, dtype=self.supk.dtype)
+        self.padding = (supn // 2, supk // 2)
+        # Set up learnable parameters
+        # for param in (rates, scales):
+        #    assert (not param) or len(param) == nkern
+        if not rates:
+
+            rates = torch.rand(nkern) * math.pi / 2.0
+
+        if not scales:
+
+            scales = (torch.rand(nkern) * 2.0 - 1.0) * math.pi / 2.0
+
+        self.rates_ = nn.Parameter(torch.Tensor(rates))
+        self.scales_ = nn.Parameter(torch.Tensor(scales))
+
+    def strfs(self):
+        """Make STRFs using the current parameters."""
+
+        if self.supn.device != self.rates_.device:  # for first run
+            self.supn = self.supn.to(self.rates_.device)
+            self.supk = self.supk.to(self.rates_.device)
+        n0, k0 = self.padding
+
+        nwind = .5 - .5 * \
+            torch.cos(2 * math.pi * (self.supn + 1) / (len(self.supn) + 1))
+        kwind = .5 - .5 * \
+            torch.cos(2 * math.pi * (self.supk + 1) / (len(self.supk) + 1))
+
+        new_wind = torch.matmul((nwind).unsqueeze(-1), (kwind).unsqueeze(0))
+
+        n_n_0 = self.supn - n0
+        k_k_0 = self.supk - k0
+        n_mult = torch.matmul(
+            n_n_0.unsqueeze(1),
+            torch.ones((1, len(self.supk))).type(torch.FloatTensor).to(
+                self.rates_.device))
+        k_mult = torch.matmul(
+            torch.ones((len(self.supn),
+                        1)).type(torch.FloatTensor).to(self.rates_.device),
+            k_k_0.unsqueeze(0))
+
+        inside = self.rates_.unsqueeze(1).unsqueeze(
+            1) * n_mult + self.scales_.unsqueeze(1).unsqueeze(1) * k_mult
+        real_strf = torch.cos(inside) * new_wind.unsqueeze(0)
+
+        if self.real_only:
+            final_strf = real_strf
+
+        else:
+            imag_strf = torch.sin(inside) * new_wind.unsqueeze(0)
+            final_strf = torch.cat([real_strf, imag_strf], dim=0)
+
+        if self.norm_strf:
+            final_strf = final_strf / (torch.sum(
+                final_strf**2, dim=(1, 2)).unsqueeze(1).unsqueeze(2))**.5
+
+        return final_strf
+
+    def forward(self, sigspec):
+        """Forward pass a batch of (real) spectra [Batch x Time x Frequency]."""
+        if len(sigspec.shape) == 2:  # expand batch dimension if single eg
+            sigspec = sigspec.unsqueeze(0)
+        strfs = self.strfs().unsqueeze(1).type_as(sigspec)
+        out = F.conv2d(sigspec.unsqueeze(1), strfs, padding=self.padding)
+        return out
+
+    def __repr__(self):
+        """Gabor filter"""
+        report = """
+            +++++ Gabor Filter Kernels [{}], supn[{}], supk[{}] real only [{}] norm strf [{}] +++++
+
+        """.format(self.numKern, self.numN, self.numK, self.real_only,
+                   self.norm_strf)
+
+        return report

modelscope/models/audio/network/se_net.py

@@ -0,0 +1,483 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..layers.activations import RectifiedLinear, Sigmoid
+from ..layers.affine_transform import AffineTransform
+from ..layers.deep_fsmn import DeepFsmn
+from ..layers.uni_deep_fsmn import Conv2d, UniDeepFsmn
+
+
+class MaskNet(nn.Module):
+
+    def __init__(self,
+                 indim,
+                 outdim,
+                 layers=9,
+                 hidden_dim=128,
+                 hidden_dim2=None,
+                 lorder=20,
+                 rorder=0,
+                 dilation=1,
+                 layer_norm=False,
+                 dropout=0,
+                 crm=False,
+                 vad=False,
+                 linearout=False):
+        super(MaskNet, self).__init__()
+
+        self.linear1 = AffineTransform(indim, hidden_dim)
+        self.relu = RectifiedLinear(hidden_dim, hidden_dim)
+        if hidden_dim2 is None:
+            hidden_dim2 = hidden_dim
+
+        if rorder == 0:
+            repeats = [
+                UniDeepFsmn(
+                    hidden_dim,
+                    hidden_dim,
+                    lorder,
+                    hidden_dim2,
+                    dilation=dilation,
+                    layer_norm=layer_norm,
+                    dropout=dropout) for i in range(layers)
+            ]
+        else:
+            repeats = [
+                DeepFsmn(
+                    hidden_dim,
+                    hidden_dim,
+                    lorder,
+                    rorder,
+                    hidden_dim2,
+                    layer_norm=layer_norm,
+                    dropout=dropout) for i in range(layers)
+            ]
+        self.deepfsmn = nn.Sequential(*repeats)
+
+        self.linear2 = AffineTransform(hidden_dim, outdim)
+
+        self.crm = crm
+        if self.crm:
+            self.sig = nn.Tanh()
+        else:
+            self.sig = Sigmoid(outdim, outdim)
+
+        self.vad = vad
+        if self.vad:
+            self.linear3 = AffineTransform(hidden_dim, 1)
+
+        self.layers = layers
+        self.linearout = linearout
+        if self.linearout and self.vad:
+            print('Warning: not supported nnet')
+
+    def forward(self, feat, ctl=None):
+        x1 = self.linear1(feat)
+        x2 = self.relu(x1)
+        if ctl is not None:
+            ctl = min(ctl, self.layers - 1)
+            for i in range(ctl):
+                x2 = self.deepfsmn[i](x2)
+            mask = self.sig(self.linear2(x2))
+            if self.vad:
+                vad = torch.sigmoid(self.linear3(x2))
+                return mask, vad
+            else:
+                return mask
+        x3 = self.deepfsmn(x2)
+        if self.linearout:
+            return self.linear2(x3)
+        mask = self.sig(self.linear2(x3))
+        if self.vad:
+            vad = torch.sigmoid(self.linear3(x3))
+            return mask, vad
+        else:
+            return mask
+
+    def to_kaldi_nnet(self):
+        re_str = ''
+        re_str += '<Nnet>\n'
+        re_str += self.linear1.to_kaldi_nnet()
+        re_str += self.relu.to_kaldi_nnet()
+        for dfsmn in self.deepfsmn:
+            re_str += dfsmn.to_kaldi_nnet()
+        re_str += self.linear2.to_kaldi_nnet()
+        re_str += self.sig.to_kaldi_nnet()
+        re_str += '</Nnet>\n'
+
+        return re_str
+
+    def to_raw_nnet(self, fid):
+        self.linear1.to_raw_nnet(fid)
+        for dfsmn in self.deepfsmn:
+            dfsmn.to_raw_nnet(fid)
+        self.linear2.to_raw_nnet(fid)
+
+
+class StageNet(nn.Module):
+
+    def __init__(self,
+                 indim,
+                 outdim,
+                 layers=9,
+                 layers2=6,
+                 hidden_dim=128,
+                 lorder=20,
+                 rorder=0,
+                 layer_norm=False,
+                 dropout=0,
+                 crm=False,
+                 vad=False,
+                 linearout=False):
+        super(StageNet, self).__init__()
+
+        self.stage1 = nn.ModuleList()
+        self.stage2 = nn.ModuleList()
+        layer = nn.Sequential(nn.Linear(indim, hidden_dim), nn.ReLU())
+        self.stage1.append(layer)
+        for i in range(layers):
+            layer = UniDeepFsmn(
+                hidden_dim,
+                hidden_dim,
+                lorder,
+                hidden_dim,
+                layer_norm=layer_norm,
+                dropout=dropout)
+            self.stage1.append(layer)
+        layer = nn.Sequential(nn.Linear(hidden_dim, 321), nn.Sigmoid())
+        self.stage1.append(layer)
+        # stage2
+        layer = nn.Sequential(nn.Linear(321 + indim, hidden_dim), nn.ReLU())
+        self.stage2.append(layer)
+        for i in range(layers2):
+            layer = UniDeepFsmn(
+                hidden_dim,
+                hidden_dim,
+                lorder,
+                hidden_dim,
+                layer_norm=layer_norm,
+                dropout=dropout)
+            self.stage2.append(layer)
+        layer = nn.Sequential(
+            nn.Linear(hidden_dim, outdim),
+            nn.Sigmoid() if not crm else nn.Tanh())
+        self.stage2.append(layer)
+        self.crm = crm
+        self.vad = vad
+        self.linearout = linearout
+        self.window = torch.hamming_window(640, periodic=False).cuda()
+        self.freezed = False
+
+    def freeze(self):
+        if not self.freezed:
+            for param in self.stage1.parameters():
+                param.requires_grad = False
+            self.freezed = True
+            print('freezed stage1')
+
+    def forward(self, feat, mixture, ctl=None):
+        if ctl == 'off':
+            x = feat
+            for i in range(len(self.stage1)):
+                x = self.stage1[i](x)
+            return x
+        else:
+            self.freeze()
+            x = feat
+            for i in range(len(self.stage1)):
+                x = self.stage1[i](x)
+
+            spec = torch.stft(
+                mixture / 32768,
+                640,
+                320,
+                640,
+                self.window,
+                center=False,
+                return_complex=True)
+            spec = torch.view_as_real(spec).permute([0, 2, 1, 3])
+            specmag = torch.sqrt(spec[..., 0]**2 + spec[..., 1]**2)
+            est = x * specmag
+            y = torch.cat([est, feat], dim=-1)
+            for i in range(len(self.stage2)):
+                y = self.stage2[i](y)
+            return y
+
+
+class Unet(nn.Module):
+
+    def __init__(self,
+                 indim,
+                 outdim,
+                 layers=9,
+                 dims=[256] * 4,
+                 lorder=20,
+                 rorder=0,
+                 dilation=1,
+                 layer_norm=False,
+                 dropout=0,
+                 crm=False,
+                 vad=False,
+                 linearout=False):
+        super(Unet, self).__init__()
+
+        self.linear1 = AffineTransform(indim, dims[0])
+        self.relu = RectifiedLinear(dims[0], dims[0])
+
+        self.encoder = nn.ModuleList()
+        self.decoder = nn.ModuleList()
+        for i in range(len(dims) - 1):
+            layer = nn.Sequential(
+                nn.Linear(dims[i], dims[i + 1]), nn.ReLU(),
+                nn.Linear(dims[i + 1], dims[i + 1], bias=False),
+                Conv2d(
+                    dims[i + 1],
+                    dims[i + 1],
+                    lorder,
+                    groups=dims[i + 1],
+                    skip_connect=True))
+            self.encoder.append(layer)
+        for i in range(len(dims) - 1, 0, -1):
+            layer = nn.Sequential(
+                nn.Linear(dims[i] * 2, dims[i - 1]), nn.ReLU(),
+                nn.Linear(dims[i - 1], dims[i - 1], bias=False),
+                Conv2d(
+                    dims[i - 1],
+                    dims[i - 1],
+                    lorder,
+                    groups=dims[i - 1],
+                    skip_connect=True))
+            self.decoder.append(layer)
+        self.tf = nn.ModuleList()
+        for i in range(layers - 2 * (len(dims) - 1)):
+            layer = nn.Sequential(
+                nn.Linear(dims[-1], dims[-1]), nn.ReLU(),
+                nn.Linear(dims[-1], dims[-1], bias=False),
+                Conv2d(
+                    dims[-1],
+                    dims[-1],
+                    lorder,
+                    groups=dims[-1],
+                    skip_connect=True))
+            self.tf.append(layer)
+
+        self.linear2 = AffineTransform(dims[0], outdim)
+        self.crm = crm
+        self.act = nn.Tanh() if self.crm else nn.Sigmoid()
+        self.vad = False
+        self.layers = layers
+        self.linearout = linearout
+
+    def forward(self, x, ctl=None):
+        x = self.linear1(x)
+        x = self.relu(x)
+
+        encoder_out = []
+        for i in range(len(self.encoder)):
+            x = self.encoder[i](x)
+            encoder_out.append(x)
+        for i in range(len(self.tf)):
+            x = self.tf[i](x)
+        for i in range(len(self.decoder)):
+            x = torch.cat([x, encoder_out[-1 - i]], dim=-1)
+            x = self.decoder[i](x)
+
+        x = self.linear2(x)
+        if self.linearout:
+            return x
+        return self.act(x)
+
+
+class BranchNet(nn.Module):
+
+    def __init__(self,
+                 indim,
+                 outdim,
+                 layers=9,
+                 hidden_dim=256,
+                 lorder=20,
+                 rorder=0,
+                 dilation=1,
+                 layer_norm=False,
+                 dropout=0,
+                 crm=False,
+                 vad=False,
+                 linearout=False):
+        super(BranchNet, self).__init__()
+
+        self.linear1 = AffineTransform(indim, hidden_dim)
+        self.relu = RectifiedLinear(hidden_dim, hidden_dim)
+
+        self.convs = nn.ModuleList()
+        self.deepfsmn = nn.ModuleList()
+        self.FREQ = nn.ModuleList()
+        self.TIME = nn.ModuleList()
+        self.br1 = nn.ModuleList()
+        self.br2 = nn.ModuleList()
+        for i in range(layers):
+            '''
+            layer = nn.Sequential(
+                nn.Linear(hidden_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, hidden_dim, bias=False),
+                Conv2d(hidden_dim, hidden_dim, lorder,
+                       groups=hidden_dim, skip_connect=True)
+            )
+            self.deepfsmn.append(layer)
+            '''
+            layer = nn.Sequential(nn.Linear(hidden_dim, hidden_dim), nn.ReLU())
+            self.FREQ.append(layer)
+            '''
+            layer = nn.GRU(hidden_dim, hidden_dim,
+                           batch_first=True,
+                           bidirectional=False)
+            self.TIME.append(layer)
+
+            layer = nn.Sequential(
+                nn.Linear(hidden_dim, hidden_dim//2, bias=False),
+                Conv2d(hidden_dim//2, hidden_dim//2, lorder,
+                       groups=hidden_dim//2, skip_connect=True)
+            )
+            self.br1.append(layer)
+            layer = nn.GRU(hidden_dim, hidden_dim//2,
+                           batch_first=True,
+                           bidirectional=False)
+            self.br2.append(layer)
+            '''
+
+        self.linear2 = AffineTransform(hidden_dim, outdim)
+        self.crm = crm
+        self.act = nn.Tanh() if self.crm else nn.Sigmoid()
+        self.vad = False
+        self.layers = layers
+        self.linearout = linearout
+
+    def forward(self, x, ctl=None):
+        return self.forward_branch(x)
+
+    def forward_sepconv(self, x):
+        x = torch.unsqueeze(x, 1)
+        for i in range(len(self.convs)):
+            x = self.convs[i](x)
+            x = F.relu(x)
+        B, C, H, W = x.shape
+        x = x.permute(0, 2, 1, 3)
+        x = torch.reshape(x, [B, H, C * W])
+        x = self.linear1(x)
+        x = self.relu(x)
+        for i in range(self.layers):
+            x = self.deepfsmn[i](x) + x
+        x = self.linear2(x)
+        return self.act(x)
+
+    def forward_branch(self, x):
+        x = self.linear1(x)
+        x = self.relu(x)
+        for i in range(self.layers):
+            z = self.FREQ[i](x)
+            x = z + x
+        x = self.linear2(x)
+        if self.linearout:
+            return x
+        return self.act(x)
+
+
+class TACNet(nn.Module):
+    ''' transform average concatenate for ad hoc dr
+    '''
+
+    def __init__(self,
+                 indim,
+                 outdim,
+                 layers=9,
+                 hidden_dim=128,
+                 lorder=20,
+                 rorder=0,
+                 crm=False,
+                 vad=False,
+                 linearout=False):
+        super(TACNet, self).__init__()
+
+        self.linear1 = AffineTransform(indim, hidden_dim)
+        self.relu = RectifiedLinear(hidden_dim, hidden_dim)
+
+        if rorder == 0:
+            repeats = [
+                UniDeepFsmn(hidden_dim, hidden_dim, lorder, hidden_dim)
+                for i in range(layers)
+            ]
+        else:
+            repeats = [
+                DeepFsmn(hidden_dim, hidden_dim, lorder, rorder, hidden_dim)
+                for i in range(layers)
+            ]
+        self.deepfsmn = nn.Sequential(*repeats)
+
+        self.ch_transform = nn.ModuleList([])
+        self.ch_average = nn.ModuleList([])
+        self.ch_concat = nn.ModuleList([])
+        for i in range(layers):
+            self.ch_transform.append(
+                nn.Sequential(nn.Linear(hidden_dim, hidden_dim), nn.PReLU()))
+            self.ch_average.append(
+                nn.Sequential(nn.Linear(hidden_dim, hidden_dim), nn.PReLU()))
+            self.ch_concat.append(
+                nn.Sequential(
+                    nn.Linear(hidden_dim * 2, hidden_dim), nn.PReLU()))
+
+        self.linear2 = AffineTransform(hidden_dim, outdim)
+
+        self.crm = crm
+        if self.crm:
+            self.sig = nn.Tanh()
+        else:
+            self.sig = Sigmoid(outdim, outdim)
+
+        self.vad = vad
+        if self.vad:
+            self.linear3 = AffineTransform(hidden_dim, 1)
+
+        self.layers = layers
+        self.linearout = linearout
+        if self.linearout and self.vad:
+            print('Warning: not supported nnet')
+
+    def forward(self, feat, ctl=None):
+        B, T, F = feat.shape
+        # assume 4ch
+        ch = 4
+        zlist = []
+        for c in range(ch):
+            z = self.linear1(feat[..., c * (F // 4):(c + 1) * (F // 4)])
+            z = self.relu(z)
+            zlist.append(z)
+        for i in range(self.layers):
+            # forward
+            for c in range(ch):
+                zlist[c] = self.deepfsmn[i](zlist[c])
+
+            # transform
+            olist = []
+            for c in range(ch):
+                z = self.ch_transform[i](zlist[c])
+                olist.append(z)
+            # average
+            avg = 0
+            for c in range(ch):
+                avg = avg + olist[c]
+            avg = avg / ch
+            avg = self.ch_average[i](avg)
+            # concate
+            for c in range(ch):
+                tac = torch.cat([olist[c], avg], dim=-1)
+                tac = self.ch_concat[i](tac)
+                zlist[c] = zlist[c] + tac
+
+        for c in range(ch):
+            zlist[c] = self.sig(self.linear2(zlist[c]))
+        mask = torch.cat(zlist, dim=-1)
+        return mask
+
+    def to_kaldi_nnet(self):
+        pass

modelscope/models/audio/tts/am/__init__.py

@@ -0,0 +1 @@
+from .sambert_hifi_16k import *  # noqa F403

modelscope/models/audio/tts/am/models/__init__.py

@@ -0,0 +1,8 @@
+from .robutrans import RobuTrans
+
+
+def create_model(name, hparams):
+    if name == 'robutrans':
+        return RobuTrans(hparams)
+    else:
+        raise Exception('Unknown model: ' + name)

modelscope/models/audio/tts/am/models/compat.py

@@ -0,0 +1,82 @@
+"""Functions for compatibility with different TensorFlow versions."""
+
+import tensorflow as tf
+
+
+def is_tf2():
+    """Returns ``True`` if running TensorFlow 2.0."""
+    return tf.__version__.startswith('2')
+
+
+def tf_supports(symbol):
+    """Returns ``True`` if TensorFlow defines :obj:`symbol`."""
+    return _string_to_tf_symbol(symbol) is not None
+
+
+def tf_any(*symbols):
+    """Returns the first supported symbol."""
+    for symbol in symbols:
+        module = _string_to_tf_symbol(symbol)
+        if module is not None:
+            return module
+    return None
+
+
+def tf_compat(v2=None, v1=None):  # pylint: disable=invalid-name
+    """Returns the compatible symbol based on the current TensorFlow version.
+
+    Args:
+      v2: The candidate v2 symbol name.
+      v1: The candidate v1 symbol name.
+
+    Returns:
+      A TensorFlow symbol.
+
+    Raises:
+      ValueError: if no symbol can be found.
+    """
+    candidates = []
+    if v2 is not None:
+        candidates.append(v2)
+    if v1 is not None:
+        candidates.append(v1)
+        candidates.append('compat.v1.%s' % v1)
+    symbol = tf_any(*candidates)
+    if symbol is None:
+        raise ValueError('Failure to resolve the TensorFlow symbol')
+    return symbol
+
+
+def name_from_variable_scope(name=''):
+    """Creates a name prefixed by the current variable scope."""
+    var_scope = tf_compat(v1='get_variable_scope')().name
+    compat_name = ''
+    if name:
+        compat_name = '%s/' % name
+    if var_scope:
+        compat_name = '%s/%s' % (var_scope, compat_name)
+    return compat_name
+
+
+def reuse():
+    """Returns ``True`` if the current variable scope is marked for reuse."""
+    return tf_compat(v1='get_variable_scope')().reuse
+
+
+def _string_to_tf_symbol(symbol):
+    modules = symbol.split('.')
+    namespace = tf
+    for module in modules:
+        namespace = getattr(namespace, module, None)
+        if namespace is None:
+            return None
+    return namespace
+
+
+# pylint: disable=invalid-name
+gfile_copy = tf_compat(v2='io.gfile.copy', v1='gfile.Copy')
+gfile_exists = tf_compat(v2='io.gfile.exists', v1='gfile.Exists')
+gfile_open = tf_compat(v2='io.gfile.GFile', v1='gfile.GFile')
+is_tensor = tf_compat(v2='is_tensor', v1='contrib.framework.is_tensor')
+logging = tf_compat(v1='logging')
+nest = tf_compat(v2='nest', v1='contrib.framework.nest')

modelscope/models/audio/tts/am/models/fsmn.py

@@ -0,0 +1,273 @@
+import tensorflow as tf
+
+
+def build_sequence_mask(sequence_length,
+                        maximum_length=None,
+                        dtype=tf.float32):
+    """Builds the dot product mask.
+
+    Args:
+      sequence_length: The sequence length.
+      maximum_length: Optional size of the returned time dimension. Otherwise
+        it is the maximum of :obj:`sequence_length`.
+      dtype: The type of the mask tensor.
+
+    Returns:
+      A broadcastable ``tf.Tensor`` of type :obj:`dtype` and shape
+      ``[batch_size, max_length]``.
+    """
+    mask = tf.sequence_mask(
+        sequence_length, maxlen=maximum_length, dtype=dtype)
+
+    return mask
+
+
+def norm(inputs):
+    """Layer normalizes :obj:`inputs`."""
+    return tf.contrib.layers.layer_norm(inputs, begin_norm_axis=-1)
+
+
+def pad_in_time(x, padding_shape):
+    """Helper function to pad a tensor in the time dimension and retain the static depth dimension.
+
+       Agrs:
+        x: [Batch, Time, Frequency]
+        padding_length: padding size of constant value (0) before the time dimension
+
+      return:
+        padded x
+    """
+
+    depth = x.get_shape().as_list()[-1]
+    x = tf.pad(x, [[0, 0], padding_shape, [0, 0]])
+    x.set_shape((None, None, depth))
+
+    return x
+
+
+def pad_in_time_right(x, padding_length):
+    """Helper function to pad a tensor in the time dimension and retain the static depth dimension.
+
+       Agrs:
+        x: [Batch, Time, Frequency]
+        padding_length: padding size of constant value (0) before the time dimension
+
+      return:
+        padded x
+    """
+    depth = x.get_shape().as_list()[-1]
+    x = tf.pad(x, [[0, 0], [0, padding_length], [0, 0]])
+    x.set_shape((None, None, depth))
+
+    return x
+
+
+def feed_forward(x, ffn_dim, memory_units, mode, dropout=0.0):
+    """Implements the Transformer's "Feed Forward" layer.
+
+    .. math::
+
+        ffn(x) = max(0, x*W_1 + b_1)*W_2
+
+    Args:
+      x: The input.
+      ffn_dim: The number of units of the nonlinear transformation.
+      memory_units: the number of units of linear transformation
+      mode: A ``tf.estimator.ModeKeys`` mode.
+      dropout: The probability to drop units from the inner transformation.
+
+    Returns:
+      The transformed input.
+    """
+    inner = tf.layers.conv1d(x, ffn_dim, 1, activation=tf.nn.relu)
+    inner = tf.layers.dropout(
+        inner, rate=dropout, training=mode == tf.estimator.ModeKeys.TRAIN)
+    outer = tf.layers.conv1d(inner, memory_units, 1, use_bias=False)
+
+    return outer
+
+
+def drop_and_add(inputs, outputs, mode, dropout=0.0):
+    """Drops units in the outputs and adds the previous values.
+
+    Args:
+      inputs: The input of the previous layer.
+      outputs: The output of the previous layer.
+      mode: A ``tf.estimator.ModeKeys`` mode.
+      dropout: The probability to drop units in :obj:`outputs`.
+
+    Returns:
+      The residual and normalized output.
+    """
+    outputs = tf.layers.dropout(outputs, rate=dropout, training=mode)
+
+    input_dim = inputs.get_shape().as_list()[-1]
+    output_dim = outputs.get_shape().as_list()[-1]
+
+    if input_dim == output_dim:
+        outputs += inputs
+
+    return outputs
+
+
+def MemoryBlock(
+    inputs,
+    filter_size,
+    mode,
+    mask=None,
+    dropout=0.0,
+):
+    """
+    Define the bidirectional memory block in FSMN
+
+    Agrs:
+      inputs: The output of the previous layer. [Batch, Time, Frequency]
+      filter_size: memory block filter size
+      mode: Training or Evaluation
+      mask: A ``tf.Tensor`` applied to the memory block output
+
+    return:
+      output: 3-D tensor ([Batch, Time, Frequency])
+    """
+    static_shape = inputs.get_shape().as_list()
+    depth = static_shape[-1]
+    inputs = tf.expand_dims(inputs, axis=1)  # [Batch, 1, Time, Frequency]
+    depthwise_filter = tf.get_variable(
+        'depth_conv_w',
+        shape=[1, filter_size, depth, 1],
+        initializer=tf.glorot_uniform_initializer(),
+        dtype=tf.float32)
+    memory = tf.nn.depthwise_conv2d(
+        input=inputs,
+        filter=depthwise_filter,
+        strides=[1, 1, 1, 1],
+        padding='SAME',
+        rate=[1, 1],
+        data_format='NHWC')
+    memory = memory + inputs
+    output = tf.layers.dropout(memory, rate=dropout, training=mode)
+    output = tf.reshape(
+        output,
+        [tf.shape(output)[0], tf.shape(output)[2], depth])
+    if mask is not None:
+        output = output * tf.expand_dims(mask, -1)
+
+    return output
+
+
+def MemoryBlockV2(
+    inputs,
+    filter_size,
+    mode,
+    shift=0,
+    mask=None,
+    dropout=0.0,
+):
+    """
+    Define the bidirectional memory block in FSMN
+
+    Agrs:
+      inputs: The output of the previous layer. [Batch, Time, Frequency]
+      filter_size: memory block filter size
+      mode: Training or Evaluation
+      shift: left padding, to control delay
+      mask: A ``tf.Tensor`` applied to the memory block output
+
+    return:
+      output: 3-D tensor ([Batch, Time, Frequency])
+    """
+    if mask is not None:
+        inputs = inputs * tf.expand_dims(mask, -1)
+
+    static_shape = inputs.get_shape().as_list()
+    depth = static_shape[-1]
+    # padding
+    left_padding = int(round((filter_size - 1) / 2))
+    right_padding = int((filter_size - 1) / 2)
+    if shift > 0:
+        left_padding = left_padding + shift
+        right_padding = right_padding - shift
+    pad_inputs = pad_in_time(inputs, [left_padding, right_padding])
+    pad_inputs = tf.expand_dims(
+        pad_inputs, axis=1)  # [Batch, 1, Time, Frequency]
+    depthwise_filter = tf.get_variable(
+        'depth_conv_w',
+        shape=[1, filter_size, depth, 1],
+        initializer=tf.glorot_uniform_initializer(),
+        dtype=tf.float32)
+    memory = tf.nn.depthwise_conv2d(
+        input=pad_inputs,
+        filter=depthwise_filter,
+        strides=[1, 1, 1, 1],
+        padding='VALID',
+        rate=[1, 1],
+        data_format='NHWC')
+    memory = tf.reshape(
+        memory,
+        [tf.shape(memory)[0], tf.shape(memory)[2], depth])
+    memory = memory + inputs
+    output = tf.layers.dropout(memory, rate=dropout, training=mode)
+    if mask is not None:
+        output = output * tf.expand_dims(mask, -1)
+
+    return output
+
+
+def UniMemoryBlock(
+    inputs,
+    filter_size,
+    mode,
+    cache=None,
+    mask=None,
+    dropout=0.0,
+):
+    """
+    Define the unidirectional memory block in FSMN
+
+    Agrs:
+      inputs: The output of the previous layer. [Batch, Time, Frequency]
+      filter_size: memory block filter size
+      cache: for streaming inference
+      mode: Training or Evaluation
+      mask: A ``tf.Tensor`` applied to the memory block output
+      dropout: dorpout factor
+    return:
+      output: 3-D tensor ([Batch, Time, Frequency])
+    """
+    if cache is not None:
+        static_shape = cache['queries'].get_shape().as_list()
+        depth = static_shape[-1]
+        queries = tf.slice(cache['queries'], [0, 1, 0], [
+            tf.shape(cache['queries'])[0],
+            tf.shape(cache['queries'])[1] - 1, depth
+        ])
+        queries = tf.concat([queries, inputs], axis=1)
+        cache['queries'] = queries
+    else:
+        padding_length = filter_size - 1
+        queries = pad_in_time(inputs, [padding_length, 0])
+
+    queries = tf.expand_dims(queries, axis=1)  # [Batch, 1, Time, Frequency]
+    static_shape = queries.get_shape().as_list()
+    depth = static_shape[-1]
+    depthwise_filter = tf.get_variable(
+        'depth_conv_w',
+        shape=[1, filter_size, depth, 1],
+        initializer=tf.glorot_uniform_initializer(),
+        dtype=tf.float32)
+    memory = tf.nn.depthwise_conv2d(
+        input=queries,
+        filter=depthwise_filter,
+        strides=[1, 1, 1, 1],
+        padding='VALID',
+        rate=[1, 1],
+        data_format='NHWC')
+    memory = tf.reshape(
+        memory,
+        [tf.shape(memory)[0], tf.shape(memory)[2], depth])
+    memory = memory + inputs
+    output = tf.layers.dropout(memory, rate=dropout, training=mode)
+    if mask is not None:
+        output = output * tf.expand_dims(mask, -1)
+
+    return output

modelscope/models/audio/tts/am/models/fsmn_encoder.py

@@ -0,0 +1,178 @@
+import tensorflow as tf
+
+from . import fsmn
+
+
+class FsmnEncoder():
+    """Encoder using Fsmn
+    """
+
+    def __init__(self,
+                 filter_size,
+                 fsmn_num_layers,
+                 dnn_num_layers,
+                 num_memory_units=512,
+                 ffn_inner_dim=2048,
+                 dropout=0.0,
+                 position_encoder=None):
+        """Initializes the parameters of the encoder.
+
+        Args:
+          filter_size: the total order of memory block
+          fsmn_num_layers: The number of fsmn layers.
+          dnn_num_layers: The number of dnn layers
+          num_units: The number of memory units.
+          ffn_inner_dim: The number of units of the inner linear transformation
+            in the feed forward layer.
+          dropout: The probability to drop units from the outputs.
+          position_encoder: The :class:`opennmt.layers.position.PositionEncoder` to
+            apply on inputs or ``None``.
+        """
+        super(FsmnEncoder, self).__init__()
+        self.filter_size = filter_size
+        self.fsmn_num_layers = fsmn_num_layers
+        self.dnn_num_layers = dnn_num_layers
+        self.num_memory_units = num_memory_units
+        self.ffn_inner_dim = ffn_inner_dim
+        self.dropout = dropout
+        self.position_encoder = position_encoder
+
+    def encode(self, inputs, sequence_length=None, mode=True):
+        if self.position_encoder is not None:
+            inputs = self.position_encoder(inputs)
+
+        inputs = tf.layers.dropout(inputs, rate=self.dropout, training=mode)
+
+        mask = fsmn.build_sequence_mask(
+            sequence_length, maximum_length=tf.shape(inputs)[1])
+
+        state = ()
+
+        for layer in range(self.fsmn_num_layers):
+            with tf.variable_scope('fsmn_layer_{}'.format(layer)):
+                with tf.variable_scope('ffn'):
+                    context = fsmn.feed_forward(
+                        inputs,
+                        self.ffn_inner_dim,
+                        self.num_memory_units,
+                        mode,
+                        dropout=self.dropout)
+
+                with tf.variable_scope('memory'):
+                    memory = fsmn.MemoryBlock(
+                        context,
+                        self.filter_size,
+                        mode,
+                        mask=mask,
+                        dropout=self.dropout)
+
+                    memory = fsmn.drop_and_add(
+                        inputs, memory, mode, dropout=self.dropout)
+
+                inputs = memory
+                state += (tf.reduce_mean(inputs, axis=1), )
+
+        for layer in range(self.dnn_num_layers):
+            with tf.variable_scope('dnn_layer_{}'.format(layer)):
+                transformed = fsmn.feed_forward(
+                    inputs,
+                    self.ffn_inner_dim,
+                    self.num_memory_units,
+                    mode,
+                    dropout=self.dropout)
+
+                inputs = transformed
+                state += (tf.reduce_mean(inputs, axis=1), )
+
+        outputs = inputs
+        return (outputs, state, sequence_length)
+
+
+class FsmnEncoderV2():
+    """Encoder using Fsmn
+    """
+
+    def __init__(self,
+                 filter_size,
+                 fsmn_num_layers,
+                 dnn_num_layers,
+                 num_memory_units=512,
+                 ffn_inner_dim=2048,
+                 dropout=0.0,
+                 shift=0,
+                 position_encoder=None):
+        """Initializes the parameters of the encoder.
+
+        Args:
+          filter_size: the total order of memory block
+          fsmn_num_layers: The number of fsmn layers.
+          dnn_num_layers: The number of dnn layers
+          num_units: The number of memory units.
+          ffn_inner_dim: The number of units of the inner linear transformation
+            in the feed forward layer.
+          dropout: The probability to drop units from the outputs.
+          shift: left padding, to control delay
+          position_encoder: The :class:`opennmt.layers.position.PositionEncoder` to
+            apply on inputs or ``None``.
+        """
+        super(FsmnEncoderV2, self).__init__()
+        self.filter_size = filter_size
+        self.fsmn_num_layers = fsmn_num_layers
+        self.dnn_num_layers = dnn_num_layers
+        self.num_memory_units = num_memory_units
+        self.ffn_inner_dim = ffn_inner_dim
+        self.dropout = dropout
+        self.shift = shift
+        if not isinstance(shift, list):
+            self.shift = [shift for _ in range(self.fsmn_num_layers)]
+        self.position_encoder = position_encoder
+
+    def encode(self, inputs, sequence_length=None, mode=True):
+        if self.position_encoder is not None:
+            inputs = self.position_encoder(inputs)
+
+        inputs = tf.layers.dropout(inputs, rate=self.dropout, training=mode)
+
+        mask = fsmn.build_sequence_mask(
+            sequence_length, maximum_length=tf.shape(inputs)[1])
+
+        state = ()
+        for layer in range(self.fsmn_num_layers):
+            with tf.variable_scope('fsmn_layer_{}'.format(layer)):
+                with tf.variable_scope('ffn'):
+                    context = fsmn.feed_forward(
+                        inputs,
+                        self.ffn_inner_dim,
+                        self.num_memory_units,
+                        mode,
+                        dropout=self.dropout)
+
+                with tf.variable_scope('memory'):
+                    memory = fsmn.MemoryBlockV2(
+                        context,
+                        self.filter_size,
+                        mode,
+                        shift=self.shift[layer],
+                        mask=mask,
+                        dropout=self.dropout)
+
+                    memory = fsmn.drop_and_add(
+                        inputs, memory, mode, dropout=self.dropout)
+
+                inputs = memory
+                state += (tf.reduce_mean(inputs, axis=1), )
+
+        for layer in range(self.dnn_num_layers):
+            with tf.variable_scope('dnn_layer_{}'.format(layer)):
+                transformed = fsmn.feed_forward(
+                    inputs,
+                    self.ffn_inner_dim,
+                    self.num_memory_units,
+                    mode,
+                    dropout=self.dropout)
+
+                inputs = transformed
+                state += (tf.reduce_mean(inputs, axis=1), )
+
+        outputs = inputs
+        return (outputs, state, sequence_length)

modelscope/models/audio/tts/am/models/helpers.py

@@ -0,0 +1,160 @@
+import numpy as np
+import tensorflow as tf
+from tensorflow.contrib.seq2seq import Helper
+
+
+class VarTestHelper(Helper):
+
+    def __init__(self, batch_size, inputs, dim):
+        with tf.name_scope('VarTestHelper'):
+            self._batch_size = batch_size
+            self._inputs = inputs
+            self._dim = dim
+
+            num_steps = tf.shape(self._inputs)[1]
+            self._lengths = tf.tile([num_steps], [self._batch_size])
+
+            self._inputs = tf.roll(inputs, shift=-1, axis=1)
+            self._init_inputs = inputs[:, 0, :]
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def sample_ids_shape(self):
+        return tf.TensorShape([])
+
+    @property
+    def sample_ids_dtype(self):
+        return np.int32
+
+    def initialize(self, name=None):
+        return (tf.tile([False], [self._batch_size]),
+                _go_frames(self._batch_size, self._dim, self._init_inputs))
+
+    def sample(self, time, outputs, state, name=None):
+        return tf.tile([0], [self._batch_size])  # Return all 0; we ignore them
+
+    def next_inputs(self, time, outputs, state, sample_ids, name=None):
+        with tf.name_scope('VarTestHelper'):
+            finished = (time + 1 >= self._lengths)
+            next_inputs = tf.concat([outputs, self._inputs[:, time, :]],
+                                    axis=-1)
+            return (finished, next_inputs, state)
+
+
+class VarTrainingHelper(Helper):
+
+    def __init__(self, targets, inputs, dim):
+        with tf.name_scope('VarTrainingHelper'):
+            self._targets = targets  # [N, T_in, 1]
+            self._batch_size = tf.shape(inputs)[0]  # N
+            self._inputs = inputs
+            self._dim = dim
+
+            num_steps = tf.shape(self._targets)[1]
+            self._lengths = tf.tile([num_steps], [self._batch_size])
+
+            self._inputs = tf.roll(inputs, shift=-1, axis=1)
+            self._init_inputs = inputs[:, 0, :]
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def sample_ids_shape(self):
+        return tf.TensorShape([])
+
+    @property
+    def sample_ids_dtype(self):
+        return np.int32
+
+    def initialize(self, name=None):
+        return (tf.tile([False], [self._batch_size]),
+                _go_frames(self._batch_size, self._dim, self._init_inputs))
+
+    def sample(self, time, outputs, state, name=None):
+        return tf.tile([0], [self._batch_size])  # Return all 0; we ignore them
+
+    def next_inputs(self, time, outputs, state, sample_ids, name=None):
+        with tf.name_scope(name or 'VarTrainingHelper'):
+            finished = (time + 1 >= self._lengths)
+            next_inputs = tf.concat(
+                [self._targets[:, time, :], self._inputs[:, time, :]], axis=-1)
+            return (finished, next_inputs, state)
+
+
+class VarTrainingSSHelper(Helper):
+
+    def __init__(self, targets, inputs, dim, global_step, schedule_begin,
+                 alpha, decay_steps):
+        with tf.name_scope('VarTrainingSSHelper'):
+            self._targets = targets  # [N, T_in, 1]
+            self._batch_size = tf.shape(inputs)[0]  # N
+            self._inputs = inputs
+            self._dim = dim
+
+            num_steps = tf.shape(self._targets)[1]
+            self._lengths = tf.tile([num_steps], [self._batch_size])
+
+            self._inputs = tf.roll(inputs, shift=-1, axis=1)
+            self._init_inputs = inputs[:, 0, :]
+
+            # for schedule sampling
+            self._global_step = global_step
+            self._schedule_begin = schedule_begin
+            self._alpha = alpha
+            self._decay_steps = decay_steps
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def sample_ids_shape(self):
+        return tf.TensorShape([])
+
+    @property
+    def sample_ids_dtype(self):
+        return np.int32
+
+    def initialize(self, name=None):
+        self._ratio = _tf_decay(self._global_step, self._schedule_begin,
+                                self._alpha, self._decay_steps)
+        return (tf.tile([False], [self._batch_size]),
+                _go_frames(self._batch_size, self._dim, self._init_inputs))
+
+    def sample(self, time, outputs, state, name=None):
+        return tf.tile([0], [self._batch_size])  # Return all 0; we ignore them
+
+    def next_inputs(self, time, outputs, state, sample_ids, name=None):
+        with tf.name_scope(name or 'VarTrainingHelper'):
+            finished = (time + 1 >= self._lengths)
+            next_inputs_tmp = tf.cond(
+                tf.less(
+                    tf.random_uniform([], minval=0, maxval=1,
+                                      dtype=tf.float32), self._ratio),
+                lambda: self._targets[:, time, :], lambda: outputs)
+            next_inputs = tf.concat(
+                [next_inputs_tmp, self._inputs[:, time, :]], axis=-1)
+            return (finished, next_inputs, state)
+
+
+def _go_frames(batch_size, dim, init_inputs):
+    '''Returns all-zero <GO> frames for a given batch size and output dimension'''
+    return tf.concat([tf.tile([[0.0]], [batch_size, dim]), init_inputs],
+                     axis=-1)
+
+
+def _tf_decay(global_step, schedule_begin, alpha, decay_steps):
+    tfr = tf.train.exponential_decay(
+        1.0,
+        global_step=global_step - schedule_begin,
+        decay_steps=decay_steps,
+        decay_rate=alpha,
+        name='tfr_decay')
+    final_tfr = tf.cond(
+        tf.less(global_step, schedule_begin), lambda: 1.0, lambda: tfr)
+    return final_tfr

modelscope/models/audio/tts/am/models/modules.py

@@ -0,0 +1,461 @@
+import tensorflow as tf
+from tensorflow.contrib.cudnn_rnn import CudnnLSTM
+from tensorflow.contrib.cudnn_rnn.python.ops import cudnn_rnn_ops
+from tensorflow.contrib.rnn import LSTMBlockCell
+
+
+def encoder_prenet(inputs,
+                   n_conv_layers,
+                   filters,
+                   kernel_size,
+                   dense_units,
+                   is_training,
+                   mask=None,
+                   scope='encoder_prenet'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i in range(n_conv_layers):
+            x = conv1d(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+        x = tf.layers.dense(
+            x, units=dense_units, activation=None, name='dense')
+    return x
+
+
+def decoder_prenet(inputs,
+                   prenet_units,
+                   dense_units,
+                   is_training,
+                   scope='decoder_prenet'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i, units in enumerate(prenet_units):
+            x = tf.layers.dense(
+                x,
+                units=units,
+                activation=tf.nn.relu,
+                name='dense_{}'.format(i))
+            x = tf.layers.dropout(
+                x, rate=0.5, training=is_training, name='dropout_{}'.format(i))
+        x = tf.layers.dense(
+            x, units=dense_units, activation=None, name='dense')
+    return x
+
+
+def encoder(inputs,
+            input_lengths,
+            n_conv_layers,
+            filters,
+            kernel_size,
+            lstm_units,
+            is_training,
+            embedded_inputs_speaker,
+            mask=None,
+            scope='encoder'):
+    with tf.variable_scope(scope):
+        x = conv_and_lstm(
+            inputs,
+            input_lengths,
+            n_conv_layers,
+            filters,
+            kernel_size,
+            lstm_units,
+            is_training,
+            embedded_inputs_speaker,
+            mask=mask)
+    return x
+
+
+def prenet(inputs, prenet_units, is_training, scope='prenet'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i, units in enumerate(prenet_units):
+            x = tf.layers.dense(
+                x,
+                units=units,
+                activation=tf.nn.relu,
+                name='dense_{}'.format(i))
+            x = tf.layers.dropout(
+                x, rate=0.5, training=is_training, name='dropout_{}'.format(i))
+    return x
+
+
+def postnet_residual_ulstm(inputs,
+                           n_conv_layers,
+                           filters,
+                           kernel_size,
+                           lstm_units,
+                           output_units,
+                           is_training,
+                           scope='postnet_residual_ulstm'):
+    with tf.variable_scope(scope):
+        x = conv_and_ulstm(inputs, None, n_conv_layers, filters, kernel_size,
+                           lstm_units, is_training)
+        x = conv1d(
+            x,
+            output_units,
+            kernel_size,
+            is_training,
+            activation=None,
+            dropout=False,
+            scope='conv1d_{}'.format(n_conv_layers - 1))
+    return x
+
+
+def postnet_residual_lstm(inputs,
+                          n_conv_layers,
+                          filters,
+                          kernel_size,
+                          lstm_units,
+                          output_units,
+                          is_training,
+                          scope='postnet_residual_lstm'):
+    with tf.variable_scope(scope):
+        x = conv_and_lstm(inputs, None, n_conv_layers, filters, kernel_size,
+                          lstm_units, is_training)
+        x = conv1d(
+            x,
+            output_units,
+            kernel_size,
+            is_training,
+            activation=None,
+            dropout=False,
+            scope='conv1d_{}'.format(n_conv_layers - 1))
+    return x
+
+
+def postnet_linear_ulstm(inputs,
+                         n_conv_layers,
+                         filters,
+                         kernel_size,
+                         lstm_units,
+                         output_units,
+                         is_training,
+                         scope='postnet_linear'):
+    with tf.variable_scope(scope):
+        x = conv_and_ulstm(inputs, None, n_conv_layers, filters, kernel_size,
+                           lstm_units, is_training)
+        x = tf.layers.dense(x, units=output_units)
+    return x
+
+
+def postnet_linear_lstm(inputs,
+                        n_conv_layers,
+                        filters,
+                        kernel_size,
+                        lstm_units,
+                        output_units,
+                        output_lengths,
+                        is_training,
+                        embedded_inputs_speaker2,
+                        mask=None,
+                        scope='postnet_linear'):
+    with tf.variable_scope(scope):
+        x = conv_and_lstm_dec(
+            inputs,
+            output_lengths,
+            n_conv_layers,
+            filters,
+            kernel_size,
+            lstm_units,
+            is_training,
+            embedded_inputs_speaker2,
+            mask=mask)
+        x = tf.layers.dense(x, units=output_units)
+    return x
+
+
+def postnet_linear(inputs,
+                   n_conv_layers,
+                   filters,
+                   kernel_size,
+                   lstm_units,
+                   output_units,
+                   output_lengths,
+                   is_training,
+                   embedded_inputs_speaker2,
+                   mask=None,
+                   scope='postnet_linear'):
+    with tf.variable_scope(scope):
+        x = conv_dec(
+            inputs,
+            output_lengths,
+            n_conv_layers,
+            filters,
+            kernel_size,
+            lstm_units,
+            is_training,
+            embedded_inputs_speaker2,
+            mask=mask)
+    return x
+
+
+def conv_and_lstm(inputs,
+                  sequence_lengths,
+                  n_conv_layers,
+                  filters,
+                  kernel_size,
+                  lstm_units,
+                  is_training,
+                  embedded_inputs_speaker,
+                  mask=None,
+                  scope='conv_and_lstm'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i in range(n_conv_layers):
+            x = conv1d(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+
+        x = tf.concat([x, embedded_inputs_speaker], axis=2)
+
+        outputs, states = tf.nn.bidirectional_dynamic_rnn(
+            LSTMBlockCell(lstm_units),
+            LSTMBlockCell(lstm_units),
+            x,
+            sequence_length=sequence_lengths,
+            dtype=tf.float32)
+        x = tf.concat(outputs, axis=-1)
+
+    return x
+
+
+def conv_and_lstm_dec(inputs,
+                      sequence_lengths,
+                      n_conv_layers,
+                      filters,
+                      kernel_size,
+                      lstm_units,
+                      is_training,
+                      embedded_inputs_speaker2,
+                      mask=None,
+                      scope='conv_and_lstm'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i in range(n_conv_layers):
+            x = conv1d(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+
+        x = tf.concat([x, embedded_inputs_speaker2], axis=2)
+
+        outputs, states = tf.nn.bidirectional_dynamic_rnn(
+            LSTMBlockCell(lstm_units),
+            LSTMBlockCell(lstm_units),
+            x,
+            sequence_length=sequence_lengths,
+            dtype=tf.float32)
+        x = tf.concat(outputs, axis=-1)
+    return x
+
+
+def conv_dec(inputs,
+             sequence_lengths,
+             n_conv_layers,
+             filters,
+             kernel_size,
+             lstm_units,
+             is_training,
+             embedded_inputs_speaker2,
+             mask=None,
+             scope='conv_and_lstm'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i in range(n_conv_layers):
+            x = conv1d(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+        x = tf.concat([x, embedded_inputs_speaker2], axis=2)
+    return x
+
+
+def conv_and_ulstm(inputs,
+                   sequence_lengths,
+                   n_conv_layers,
+                   filters,
+                   kernel_size,
+                   lstm_units,
+                   is_training,
+                   scope='conv_and_ulstm'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i in range(n_conv_layers):
+            x = conv1d(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                scope='conv1d_{}'.format(i))
+
+        outputs, states = tf.nn.dynamic_rnn(
+            LSTMBlockCell(lstm_units),
+            x,
+            sequence_length=sequence_lengths,
+            dtype=tf.float32)
+
+    return outputs
+
+
+def conv1d(inputs,
+           filters,
+           kernel_size,
+           is_training,
+           activation=None,
+           dropout=False,
+           mask=None,
+           scope='conv1d'):
+    with tf.variable_scope(scope):
+        if mask is not None:
+            inputs = inputs * tf.expand_dims(mask, -1)
+        x = tf.layers.conv1d(
+            inputs, filters=filters, kernel_size=kernel_size, padding='same')
+        if mask is not None:
+            x = x * tf.expand_dims(mask, -1)
+
+        x = tf.layers.batch_normalization(x, training=is_training)
+        if activation is not None:
+            x = activation(x)
+        if dropout:
+            x = tf.layers.dropout(x, rate=0.5, training=is_training)
+    return x
+
+
+def conv1d_dp(inputs,
+              filters,
+              kernel_size,
+              is_training,
+              activation=None,
+              dropout=False,
+              dropoutrate=0.5,
+              mask=None,
+              scope='conv1d'):
+    with tf.variable_scope(scope):
+        if mask is not None:
+            inputs = inputs * tf.expand_dims(mask, -1)
+        x = tf.layers.conv1d(
+            inputs, filters=filters, kernel_size=kernel_size, padding='same')
+        if mask is not None:
+            x = x * tf.expand_dims(mask, -1)
+
+        x = tf.contrib.layers.layer_norm(x)
+        if activation is not None:
+            x = activation(x)
+        if dropout:
+            x = tf.layers.dropout(x, rate=dropoutrate, training=is_training)
+    return x
+
+
+def duration_predictor(inputs,
+                       n_conv_layers,
+                       filters,
+                       kernel_size,
+                       lstm_units,
+                       input_lengths,
+                       is_training,
+                       embedded_inputs_speaker,
+                       mask=None,
+                       scope='duration_predictor'):
+    with tf.variable_scope(scope):
+        x = inputs
+        for i in range(n_conv_layers):
+            x = conv1d_dp(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                dropoutrate=0.1,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+
+        x = tf.concat([x, embedded_inputs_speaker], axis=2)
+
+        outputs, states = tf.nn.bidirectional_dynamic_rnn(
+            LSTMBlockCell(lstm_units),
+            LSTMBlockCell(lstm_units),
+            x,
+            sequence_length=input_lengths,
+            dtype=tf.float32)
+        x = tf.concat(outputs, axis=-1)
+
+        x = tf.layers.dense(x, units=1)
+        x = tf.nn.relu(x)
+    return x
+
+
+def duration_predictor2(inputs,
+                        n_conv_layers,
+                        filters,
+                        kernel_size,
+                        input_lengths,
+                        is_training,
+                        mask=None,
+                        scope='duration_predictor'):
+    with tf.variable_scope(scope):
+        x = inputs
+        for i in range(n_conv_layers):
+            x = conv1d_dp(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                dropoutrate=0.1,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+
+        x = tf.layers.dense(x, units=1)
+        x = tf.nn.relu(x)
+    return x
+
+
+def conv_prenet(inputs,
+                n_conv_layers,
+                filters,
+                kernel_size,
+                is_training,
+                mask=None,
+                scope='conv_prenet'):
+    x = inputs
+    with tf.variable_scope(scope):
+        for i in range(n_conv_layers):
+            x = conv1d(
+                x,
+                filters,
+                kernel_size,
+                is_training,
+                activation=tf.nn.relu,
+                dropout=True,
+                mask=mask,
+                scope='conv1d_{}'.format(i))
+
+    return x

modelscope/models/audio/tts/am/models/position.py

@@ -0,0 +1,174 @@
+"""Define position encoder classes."""
+
+import abc
+import math
+
+import tensorflow as tf
+
+from .reducer import SumReducer
+
+
+class PositionEncoder(tf.keras.layers.Layer):
+    """Base class for position encoders."""
+
+    def __init__(self, reducer=None, **kwargs):
+        """Initializes the position encoder.
+        Args:
+          reducer: A :class:`opennmt.layers.Reducer` to merge inputs and position
+            encodings. Defaults to :class:`opennmt.layers.SumReducer`.
+          **kwargs: Additional layer keyword arguments.
+        """
+        super(PositionEncoder, self).__init__(**kwargs)
+        if reducer is None:
+            reducer = SumReducer(dtype=kwargs.get('dtype'))
+        self.reducer = reducer
+
+    def call(self, inputs, position=None):  # pylint: disable=arguments-differ
+        """Add position encodings to :obj:`inputs`.
+        Args:
+          inputs: The inputs to encode.
+          position: The single position to encode, to use when this layer is called
+            step by step.
+        Returns:
+          A ``tf.Tensor`` whose shape depends on the configured ``reducer``.
+        """
+        batch_size = tf.shape(inputs)[0]
+        timesteps = tf.shape(inputs)[1]
+        input_dim = inputs.shape[-1].value
+        positions = tf.range(timesteps) + 1 if position is None else [position]
+        position_encoding = self._encode([positions], input_dim)
+        position_encoding = tf.tile(position_encoding, [batch_size, 1, 1])
+        return self.reducer([inputs, position_encoding])
+
+    @abc.abstractmethod
+    def _encode(self, positions, depth):
+        """Creates position encodings.
+        Args:
+          positions: The positions to encode of shape :math:`[B, ...]`.
+          depth: The encoding depth :math:`D`.
+        Returns:
+          A ``tf.Tensor`` of shape :math:`[B, ..., D]`.
+        """
+        raise NotImplementedError()
+
+
+class PositionEmbedder(PositionEncoder):
+    """Encodes position with a lookup table."""
+
+    def __init__(self, maximum_position=128, reducer=None, **kwargs):
+        """Initializes the position encoder.
+        Args:
+          maximum_position: The maximum position to embed. Positions greater
+            than this value will be set to :obj:`maximum_position`.
+          reducer: A :class:`opennmt.layers.Reducer` to merge inputs and position
+            encodings. Defaults to :class:`opennmt.layers.SumReducer`.
+          **kwargs: Additional layer keyword arguments.
+        """
+        super(PositionEmbedder, self).__init__(reducer=reducer, **kwargs)
+        self.maximum_position = maximum_position
+        self.embedding = None
+
+    def build(self, input_shape):
+        shape = [self.maximum_position + 1, input_shape[-1]]
+        self.embedding = self.add_weight('position_embedding', shape)
+        super(PositionEmbedder, self).build(input_shape)
+
+    def _encode(self, positions, depth):
+        positions = tf.minimum(positions, self.maximum_position)
+        return tf.nn.embedding_lookup(self.embedding, positions)
+
+
+class SinusoidalPositionEncoder(PositionEncoder):
+    """Encodes positions with sine waves as described in
+    https://arxiv.org/abs/1706.03762.
+    """
+
+    def _encode(self, positions, depth):
+        if depth % 2 != 0:
+            raise ValueError(
+                'SinusoidalPositionEncoder expects the depth to be divisble '
+                'by 2 but got %d' % depth)
+
+        batch_size = tf.shape(positions)[0]
+        positions = tf.cast(positions, tf.float32)
+
+        log_timescale_increment = math.log(10000) / (depth / 2 - 1)
+        inv_timescales = tf.exp(
+            tf.range(depth / 2, dtype=tf.float32) * -log_timescale_increment)
+        inv_timescales = tf.reshape(
+            tf.tile(inv_timescales, [batch_size]), [batch_size, depth // 2])
+        scaled_time = tf.expand_dims(positions, -1) * tf.expand_dims(
+            inv_timescales, 1)
+        encoding = tf.concat(
+            [tf.sin(scaled_time), tf.cos(scaled_time)], axis=2)
+        return tf.cast(encoding, self.dtype)
+
+
+class SinusodalPositionalEncoding(tf.keras.layers.Layer):
+
+    def __init__(self, name='SinusodalPositionalEncoding'):
+        super(SinusodalPositionalEncoding, self).__init__(name=name)
+
+    @staticmethod
+    def positional_encoding(len, dim, step=1.):
+        """
+        :param len: int scalar
+        :param dim: int scalar
+        :param step:
+        :return: position embedding
+        """
+        pos_mat = tf.tile(
+            tf.expand_dims(
+                tf.range(0, tf.cast(len, dtype=tf.float32), dtype=tf.float32)
+                * step,
+                axis=-1), [1, dim])
+        dim_mat = tf.tile(
+            tf.expand_dims(
+                tf.range(0, tf.cast(dim, dtype=tf.float32), dtype=tf.float32),
+                axis=0), [len, 1])
+        dim_mat_int = tf.cast(dim_mat, dtype=tf.int32)
+        pos_encoding = tf.where(  # [time, dims]
+            tf.math.equal(tf.math.mod(dim_mat_int, 2), 0),
+            x=tf.math.sin(
+                pos_mat / tf.pow(10000., dim_mat / tf.cast(dim, tf.float32))),
+            y=tf.math.cos(pos_mat
+                          / tf.pow(10000.,
+                                   (dim_mat - 1) / tf.cast(dim, tf.float32))))
+        return pos_encoding
+
+
+class BatchSinusodalPositionalEncoding(tf.keras.layers.Layer):
+
+    def __init__(self, name='BatchSinusodalPositionalEncoding'):
+        super(BatchSinusodalPositionalEncoding, self).__init__(name=name)
+
+    @staticmethod
+    def positional_encoding(batch_size, len, dim, pos_mat, step=1.):
+        """
+        :param len: int scalar
+        :param dim: int scalar
+        :param step:
+        :param pos_mat: [B, len] = [len, 1] * dim
+        :return: position embedding
+        """
+        pos_mat = tf.tile(
+            tf.expand_dims(tf.cast(pos_mat, dtype=tf.float32) * step, axis=-1),
+            [1, 1, dim])  # [B, len, dim]
+
+        dim_mat = tf.tile(
+            tf.expand_dims(
+                tf.expand_dims(
+                    tf.range(
+                        0, tf.cast(dim, dtype=tf.float32), dtype=tf.float32),
+                    axis=0),
+                axis=0), [batch_size, len, 1])  # [B, len, dim]
+
+        dim_mat_int = tf.cast(dim_mat, dtype=tf.int32)
+        pos_encoding = tf.where(  # [B, time, dims]
+            tf.math.equal(tf.mod(dim_mat_int, 2), 0),
+            x=tf.math.sin(
+                pos_mat / tf.pow(10000., dim_mat / tf.cast(dim, tf.float32))),
+            y=tf.math.cos(pos_mat
+                          / tf.pow(10000.,
+                                   (dim_mat - 1) / tf.cast(dim, tf.float32))))
+        return pos_encoding

modelscope/models/audio/tts/am/models/reducer.py

@@ -0,0 +1,155 @@
+"""Define reducers: objects that merge inputs."""
+
+import abc
+import functools
+
+import tensorflow as tf
+
+
+def pad_in_time(x, padding_length):
+    """Helper function to pad a tensor in the time dimension and retain the static depth dimension."""
+    return tf.pad(x, [[0, 0], [0, padding_length], [0, 0]])
+
+
+def align_in_time(x, length):
+    """Aligns the time dimension of :obj:`x` with :obj:`length`."""
+    time_dim = tf.shape(x)[1]
+    return tf.cond(
+        tf.less(time_dim, length),
+        true_fn=lambda: pad_in_time(x, length - time_dim),
+        false_fn=lambda: x[:, :length])
+
+
+def pad_with_identity(x,
+                      sequence_length,
+                      max_sequence_length,
+                      identity_values=0,
+                      maxlen=None):
+    """Pads a tensor with identity values up to :obj:`max_sequence_length`.
+    Args:
+      x: A ``tf.Tensor`` of shape ``[batch_size, time, depth]``.
+      sequence_length: The true sequence length of :obj:`x`.
+      max_sequence_length: The sequence length up to which the tensor must contain
+        :obj:`identity values`.
+      identity_values: The identity value.
+      maxlen: Size of the output time dimension. Default is the maximum value in
+        obj:`max_sequence_length`.
+    Returns:
+      A ``tf.Tensor`` of shape ``[batch_size, maxlen, depth]``.
+    """
+    if maxlen is None:
+        maxlen = tf.reduce_max(max_sequence_length)
+
+    mask = tf.sequence_mask(sequence_length, maxlen=maxlen, dtype=x.dtype)
+    mask = tf.expand_dims(mask, axis=-1)
+    mask_combined = tf.sequence_mask(
+        max_sequence_length, maxlen=maxlen, dtype=x.dtype)
+    mask_combined = tf.expand_dims(mask_combined, axis=-1)
+
+    identity_mask = mask_combined * (1.0 - mask)
+
+    x = pad_in_time(x, maxlen - tf.shape(x)[1])
+    x = x * mask + (identity_mask * identity_values)
+
+    return x
+
+
+def pad_n_with_identity(inputs, sequence_lengths, identity_values=0):
+    """Pads each input tensors with identity values up to
+    ``max(sequence_lengths)`` for each batch.
+    Args:
+      inputs: A list of ``tf.Tensor``.
+      sequence_lengths: A list of sequence length.
+      identity_values: The identity value.
+    Returns:
+      A tuple ``(padded, max_sequence_length)`` which are respectively a list of
+      ``tf.Tensor`` where each tensor are padded with identity and the combined
+      sequence length.
+    """
+    max_sequence_length = tf.reduce_max(sequence_lengths, axis=0)
+    maxlen = tf.reduce_max([tf.shape(x)[1] for x in inputs])
+    padded = [
+        pad_with_identity(
+            x,
+            length,
+            max_sequence_length,
+            identity_values=identity_values,
+            maxlen=maxlen) for x, length in zip(inputs, sequence_lengths)
+    ]
+    return padded, max_sequence_length
+
+
+class Reducer(tf.keras.layers.Layer):
+    """Base class for reducers."""
+
+    def zip_and_reduce(self, x, y):
+        """Zips the :obj:`x` with :obj:`y` structures together and reduces all
+        elements. If the structures are nested, they will be flattened first.
+        Args:
+          x: The first structure.
+          y: The second structure.
+        Returns:
+          The same structure as :obj:`x` and :obj:`y` where each element from
+          :obj:`x` is reduced with the correspond element from :obj:`y`.
+        Raises:
+          ValueError: if the two structures are not the same.
+        """
+        tf.nest.assert_same_structure(x, y)
+        x_flat = tf.nest.flatten(x)
+        y_flat = tf.nest.flatten(y)
+        reduced = list(map(self, zip(x_flat, y_flat)))
+        return tf.nest.pack_sequence_as(x, reduced)
+
+    def call(self, inputs, sequence_length=None):  # pylint: disable=arguments-differ
+        """Reduces all input elements.
+        Args:
+          inputs: A list of ``tf.Tensor``.
+          sequence_length: The length of each input, if reducing sequences.
+        Returns:
+          If :obj:`sequence_length` is set, a tuple
+          ``(reduced_input, reduced_length)``, otherwise a reduced ``tf.Tensor``
+          only.
+        """
+        if sequence_length is None:
+            return self.reduce(inputs)
+        else:
+            return self.reduce_sequence(
+                inputs, sequence_lengths=sequence_length)
+
+    @abc.abstractmethod
+    def reduce(self, inputs):
+        """See :meth:`opennmt.layers.Reducer.__call__`."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def reduce_sequence(self, inputs, sequence_lengths):
+        """See :meth:`opennmt.layers.Reducer.__call__`."""
+        raise NotImplementedError()
+
+
+class SumReducer(Reducer):
+    """A reducer that sums the inputs."""
+
+    def reduce(self, inputs):
+        if len(inputs) == 1:
+            return inputs[0]
+        if len(inputs) == 2:
+            return inputs[0] + inputs[1]
+        return tf.add_n(inputs)
+
+    def reduce_sequence(self, inputs, sequence_lengths):
+        padded, combined_length = pad_n_with_identity(
+            inputs, sequence_lengths, identity_values=0)
+        return self.reduce(padded), combined_length
+
+
+class MultiplyReducer(Reducer):
+    """A reducer that multiplies the inputs."""
+
+    def reduce(self, inputs):
+        return functools.reduce(lambda a, x: a * x, inputs)
+
+    def reduce_sequence(self, inputs, sequence_lengths):
+        padded, combined_length = pad_n_with_identity(
+            inputs, sequence_lengths, identity_values=1)
+        return self.reduce(padded), combined_length

modelscope/models/audio/tts/am/models/rnn_wrappers.py

@@ -0,0 +1,240 @@
+import numpy as np
+import tensorflow as tf
+from tensorflow.contrib.rnn import RNNCell
+from tensorflow.contrib.seq2seq import AttentionWrapperState
+from tensorflow.python.ops import rnn_cell_impl
+
+from .modules import prenet
+
+
+class VarPredictorCell(RNNCell):
+    '''Wrapper wrapper knock knock.'''
+
+    def __init__(self, var_predictor_cell, is_training, dim, prenet_units):
+        super(VarPredictorCell, self).__init__()
+        self._var_predictor_cell = var_predictor_cell
+        self._is_training = is_training
+        self._dim = dim
+        self._prenet_units = prenet_units
+
+    @property
+    def state_size(self):
+        return tuple([self.output_size, self._var_predictor_cell.state_size])
+
+    @property
+    def output_size(self):
+        return self._dim
+
+    def zero_state(self, batch_size, dtype):
+        return tuple([
+            rnn_cell_impl._zero_state_tensors(self.output_size, batch_size,
+                                              dtype),
+            self._var_predictor_cell.zero_state(batch_size, dtype)
+        ])
+
+    def call(self, inputs, state):
+        '''Run the Tacotron2 super decoder cell.'''
+        super_cell_out, decoder_state = state
+
+        # split
+        prenet_input = inputs[:, 0:self._dim]
+        encoder_output = inputs[:, self._dim:]
+
+        # prenet and concat
+        prenet_output = prenet(
+            prenet_input,
+            self._prenet_units,
+            self._is_training,
+            scope='var_prenet')
+        decoder_input = tf.concat([prenet_output, encoder_output], axis=-1)
+
+        # decoder LSTM/GRU
+        new_super_cell_out, new_decoder_state = self._var_predictor_cell(
+            decoder_input, decoder_state)
+
+        # projection
+        new_super_cell_out = tf.layers.dense(
+            new_super_cell_out, units=self._dim)
+
+        new_states = tuple([new_super_cell_out, new_decoder_state])
+
+        return new_super_cell_out, new_states
+
+
+class DurPredictorCell(RNNCell):
+    '''Wrapper wrapper knock knock.'''
+
+    def __init__(self, var_predictor_cell, is_training, dim, prenet_units):
+        super(DurPredictorCell, self).__init__()
+        self._var_predictor_cell = var_predictor_cell
+        self._is_training = is_training
+        self._dim = dim
+        self._prenet_units = prenet_units
+
+    @property
+    def state_size(self):
+        return tuple([self.output_size, self._var_predictor_cell.state_size])
+
+    @property
+    def output_size(self):
+        return self._dim
+
+    def zero_state(self, batch_size, dtype):
+        return tuple([
+            rnn_cell_impl._zero_state_tensors(self.output_size, batch_size,
+                                              dtype),
+            self._var_predictor_cell.zero_state(batch_size, dtype)
+        ])
+
+    def call(self, inputs, state):
+        '''Run the Tacotron2 super decoder cell.'''
+        super_cell_out, decoder_state = state
+
+        # split
+        prenet_input = inputs[:, 0:self._dim]
+        encoder_output = inputs[:, self._dim:]
+
+        # prenet and concat
+        prenet_output = prenet(
+            prenet_input,
+            self._prenet_units,
+            self._is_training,
+            scope='dur_prenet')
+        decoder_input = tf.concat([prenet_output, encoder_output], axis=-1)
+
+        # decoder LSTM/GRU
+        new_super_cell_out, new_decoder_state = self._var_predictor_cell(
+            decoder_input, decoder_state)
+
+        # projection
+        new_super_cell_out = tf.layers.dense(
+            new_super_cell_out, units=self._dim)
+        new_super_cell_out = tf.nn.relu(new_super_cell_out)
+        #    new_super_cell_out = tf.log(tf.cast(tf.round(tf.exp(new_super_cell_out) - 1), tf.float32) + 1)
+
+        new_states = tuple([new_super_cell_out, new_decoder_state])
+
+        return new_super_cell_out, new_states
+
+
+class DurPredictorCECell(RNNCell):
+    '''Wrapper wrapper knock knock.'''
+
+    def __init__(self, var_predictor_cell, is_training, dim, prenet_units,
+                 max_dur, dur_embedding_dim):
+        super(DurPredictorCECell, self).__init__()
+        self._var_predictor_cell = var_predictor_cell
+        self._is_training = is_training
+        self._dim = dim
+        self._prenet_units = prenet_units
+        self._max_dur = max_dur
+        self._dur_embedding_dim = dur_embedding_dim
+
+    @property
+    def state_size(self):
+        return tuple([self.output_size, self._var_predictor_cell.state_size])
+
+    @property
+    def output_size(self):
+        return self._max_dur
+
+    def zero_state(self, batch_size, dtype):
+        return tuple([
+            rnn_cell_impl._zero_state_tensors(self.output_size, batch_size,
+                                              dtype),
+            self._var_predictor_cell.zero_state(batch_size, dtype)
+        ])
+
+    def call(self, inputs, state):
+        '''Run the Tacotron2 super decoder cell.'''
+        super_cell_out, decoder_state = state
+
+        # split
+        prenet_input = tf.squeeze(
+            tf.cast(inputs[:, 0:self._dim], tf.int32), axis=-1)  # [N]
+        prenet_input = tf.one_hot(
+            prenet_input, self._max_dur, on_value=1.0, off_value=0.0,
+            axis=-1)  # [N, 120]
+        prenet_input = tf.layers.dense(
+            prenet_input, units=self._dur_embedding_dim)
+        encoder_output = inputs[:, self._dim:]
+
+        # prenet and concat
+        prenet_output = prenet(
+            prenet_input,
+            self._prenet_units,
+            self._is_training,
+            scope='dur_prenet')
+        decoder_input = tf.concat([prenet_output, encoder_output], axis=-1)
+
+        # decoder LSTM/GRU
+        new_super_cell_out, new_decoder_state = self._var_predictor_cell(
+            decoder_input, decoder_state)
+
+        # projection
+        new_super_cell_out = tf.layers.dense(
+            new_super_cell_out, units=self._max_dur)  # [N, 120]
+        new_super_cell_out = tf.nn.softmax(new_super_cell_out)  # [N, 120]
+
+        new_states = tuple([new_super_cell_out, new_decoder_state])
+
+        return new_super_cell_out, new_states
+
+
+class VarPredictorCell2(RNNCell):
+    '''Wrapper wrapper knock knock.'''
+
+    def __init__(self, var_predictor_cell, is_training, dim, prenet_units):
+        super(VarPredictorCell2, self).__init__()
+        self._var_predictor_cell = var_predictor_cell
+        self._is_training = is_training
+        self._dim = dim
+        self._prenet_units = prenet_units
+
+    @property
+    def state_size(self):
+        return tuple([self.output_size, self._var_predictor_cell.state_size])
+
+    @property
+    def output_size(self):
+        return self._dim
+
+    def zero_state(self, batch_size, dtype):
+        return tuple([
+            rnn_cell_impl._zero_state_tensors(self.output_size, batch_size,
+                                              dtype),
+            self._var_predictor_cell.zero_state(batch_size, dtype)
+        ])
+
+    def call(self, inputs, state):
+        '''Run the Tacotron2 super decoder cell.'''
+        super_cell_out, decoder_state = state
+
+        # split
+        prenet_input = inputs[:, 0:self._dim]
+        encoder_output = inputs[:, self._dim:]
+
+        # prenet and concat
+        prenet_output = prenet(
+            prenet_input,
+            self._prenet_units,
+            self._is_training,
+            scope='var_prenet')
+        decoder_input = tf.concat([prenet_output, encoder_output], axis=-1)
+
+        # decoder LSTM/GRU
+        new_super_cell_out, new_decoder_state = self._var_predictor_cell(
+            decoder_input, decoder_state)
+
+        # projection
+        new_super_cell_out = tf.layers.dense(
+            new_super_cell_out, units=self._dim)
+
+        # split and relu
+        new_super_cell_out = tf.concat([
+            tf.nn.relu(new_super_cell_out[:, 0:1]), new_super_cell_out[:, 1:]
+        ], axis=-1)  # yapf:disable
+
+        new_states = tuple([new_super_cell_out, new_decoder_state])
+
+        return new_super_cell_out, new_states

modelscope/models/audio/tts/am/models/robutrans.py

@@ -0,0 +1,760 @@
+import tensorflow as tf
+from tensorflow.contrib.rnn import LSTMBlockCell, MultiRNNCell
+from tensorflow.contrib.seq2seq import BasicDecoder
+from tensorflow.python.ops.ragged.ragged_util import repeat
+
+from .fsmn_encoder import FsmnEncoderV2
+from .helpers import VarTestHelper, VarTrainingHelper
+from .modules import conv_prenet, decoder_prenet, encoder_prenet
+from .position import (BatchSinusodalPositionalEncoding,
+                       SinusodalPositionalEncoding)
+from .rnn_wrappers import DurPredictorCell, VarPredictorCell
+from .self_attention_decoder import SelfAttentionDecoder
+from .self_attention_encoder import SelfAttentionEncoder
+
+
+class RobuTrans():
+
+    def __init__(self, hparams):
+        self._hparams = hparams
+
+    def initialize(self,
+                   inputs,
+                   inputs_emotion,
+                   inputs_speaker,
+                   input_lengths,
+                   output_lengths=None,
+                   mel_targets=None,
+                   durations=None,
+                   pitch_contours=None,
+                   uv_masks=None,
+                   pitch_scales=None,
+                   duration_scales=None,
+                   energy_contours=None,
+                   energy_scales=None):
+        '''Initializes the model for inference.
+
+        Sets "mel_outputs", "linear_outputs", "stop_token_outputs", and "alignments" fields.
+
+        Args:
+          inputs: int32 Tensor with shape [N, T_in] where N is batch size, T_in is number of
+            steps in the input time series, and values are character IDs
+          input_lengths: int32 Tensor with shape [N] where N is batch size and values are the lengths
+            of each sequence in inputs.
+          output_lengths: int32 Tensor with shape [N] where N is batch size and values are the lengths
+            of each sequence in outputs.
+          mel_targets: float32 Tensor with shape [N, T_out, M] where N is batch size, T_out is number
+            of steps in the output time series, M is num_mels, and values are entries in the mel
+            spectrogram. Only needed for training.
+        '''
+        with tf.variable_scope('inference') as _:
+            is_training = mel_targets is not None
+            batch_size = tf.shape(inputs)[0]
+            hp = self._hparams
+
+            input_mask = None
+            if input_lengths is not None and is_training:
+                input_mask = tf.sequence_mask(
+                    input_lengths, tf.shape(inputs)[1], dtype=tf.float32)
+
+            if input_mask is not None:
+                inputs = inputs * tf.expand_dims(input_mask, -1)
+
+            # speaker embedding
+            embedded_inputs_speaker = tf.layers.dense(
+                inputs_speaker,
+                32,
+                activation=None,
+                use_bias=False,
+                kernel_initializer=tf.truncated_normal_initializer(stddev=0.5))
+
+            # emotion embedding
+            embedded_inputs_emotion = tf.layers.dense(
+                inputs_emotion,
+                32,
+                activation=None,
+                use_bias=False,
+                kernel_initializer=tf.truncated_normal_initializer(stddev=0.5))
+
+            # symbol embedding
+            with tf.variable_scope('Embedding'):
+                embedded_inputs = tf.layers.dense(
+                    inputs,
+                    hp.embedding_dim,
+                    activation=None,
+                    use_bias=False,
+                    kernel_initializer=tf.truncated_normal_initializer(
+                        stddev=0.5))
+
+            # Encoder
+            with tf.variable_scope('Encoder'):
+                Encoder = SelfAttentionEncoder(
+                    num_layers=hp.encoder_num_layers,
+                    num_units=hp.encoder_num_units,
+                    num_heads=hp.encoder_num_heads,
+                    ffn_inner_dim=hp.encoder_ffn_inner_dim,
+                    dropout=hp.encoder_dropout,
+                    attention_dropout=hp.encoder_attention_dropout,
+                    relu_dropout=hp.encoder_relu_dropout)
+                encoder_outputs, state_mo, sequence_length_mo, attns = Encoder.encode(
+                    embedded_inputs,
+                    sequence_length=input_lengths,
+                    mode=is_training)
+                encoder_outputs = tf.layers.dense(
+                    encoder_outputs,
+                    hp.encoder_projection_units,
+                    activation=None,
+                    use_bias=False,
+                    kernel_initializer=tf.truncated_normal_initializer(
+                        stddev=0.5))
+
+            # pitch and energy
+            var_inputs = tf.concat([
+                encoder_outputs, embedded_inputs_speaker,
+                embedded_inputs_emotion
+            ], 2)
+            if input_mask is not None:
+                var_inputs = var_inputs * tf.expand_dims(input_mask, -1)
+
+            with tf.variable_scope('Pitch_Predictor'):
+                Pitch_Predictor_FSMN = FsmnEncoderV2(
+                    filter_size=hp.predictor_filter_size,
+                    fsmn_num_layers=hp.predictor_fsmn_num_layers,
+                    dnn_num_layers=hp.predictor_dnn_num_layers,
+                    num_memory_units=hp.predictor_num_memory_units,
+                    ffn_inner_dim=hp.predictor_ffn_inner_dim,
+                    dropout=hp.predictor_dropout,
+                    shift=hp.predictor_shift,
+                    position_encoder=None)
+                pitch_contour_outputs, _, _ = Pitch_Predictor_FSMN.encode(
+                    tf.concat([
+                        encoder_outputs, embedded_inputs_speaker,
+                        embedded_inputs_emotion
+                    ], 2),
+                    sequence_length=input_lengths,
+                    mode=is_training)
+                pitch_contour_outputs, _ = tf.nn.bidirectional_dynamic_rnn(
+                    LSTMBlockCell(hp.predictor_lstm_units),
+                    LSTMBlockCell(hp.predictor_lstm_units),
+                    pitch_contour_outputs,
+                    sequence_length=input_lengths,
+                    dtype=tf.float32)
+                pitch_contour_outputs = tf.concat(
+                    pitch_contour_outputs, axis=-1)
+                pitch_contour_outputs = tf.layers.dense(
+                    pitch_contour_outputs, units=1)  # [N, T_in, 1]
+                pitch_contour_outputs = tf.squeeze(
+                    pitch_contour_outputs, axis=2)  # [N, T_in]
+
+            with tf.variable_scope('Energy_Predictor'):
+                Energy_Predictor_FSMN = FsmnEncoderV2(
+                    filter_size=hp.predictor_filter_size,
+                    fsmn_num_layers=hp.predictor_fsmn_num_layers,
+                    dnn_num_layers=hp.predictor_dnn_num_layers,
+                    num_memory_units=hp.predictor_num_memory_units,
+                    ffn_inner_dim=hp.predictor_ffn_inner_dim,
+                    dropout=hp.predictor_dropout,
+                    shift=hp.predictor_shift,
+                    position_encoder=None)
+                energy_contour_outputs, _, _ = Energy_Predictor_FSMN.encode(
+                    tf.concat([
+                        encoder_outputs, embedded_inputs_speaker,
+                        embedded_inputs_emotion
+                    ], 2),
+                    sequence_length=input_lengths,
+                    mode=is_training)
+                energy_contour_outputs, _ = tf.nn.bidirectional_dynamic_rnn(
+                    LSTMBlockCell(hp.predictor_lstm_units),
+                    LSTMBlockCell(hp.predictor_lstm_units),
+                    energy_contour_outputs,
+                    sequence_length=input_lengths,
+                    dtype=tf.float32)
+                energy_contour_outputs = tf.concat(
+                    energy_contour_outputs, axis=-1)
+                energy_contour_outputs = tf.layers.dense(
+                    energy_contour_outputs, units=1)  # [N, T_in, 1]
+                energy_contour_outputs = tf.squeeze(
+                    energy_contour_outputs, axis=2)  # [N, T_in]
+
+            if is_training:
+                pitch_embeddings = tf.expand_dims(
+                    pitch_contours, axis=2)  # [N, T_in, 1]
+                pitch_embeddings = tf.layers.conv1d(
+                    pitch_embeddings,
+                    filters=hp.encoder_projection_units,
+                    kernel_size=9,
+                    padding='same',
+                    name='pitch_embeddings')  # [N, T_in, 32]
+
+                energy_embeddings = tf.expand_dims(
+                    energy_contours, axis=2)  # [N, T_in, 1]
+                energy_embeddings = tf.layers.conv1d(
+                    energy_embeddings,
+                    filters=hp.encoder_projection_units,
+                    kernel_size=9,
+                    padding='same',
+                    name='energy_embeddings')  # [N, T_in, 32]
+            else:
+                pitch_contour_outputs *= pitch_scales
+                pitch_embeddings = tf.expand_dims(
+                    pitch_contour_outputs, axis=2)  # [N, T_in, 1]
+                pitch_embeddings = tf.layers.conv1d(
+                    pitch_embeddings,
+                    filters=hp.encoder_projection_units,
+                    kernel_size=9,
+                    padding='same',
+                    name='pitch_embeddings')  # [N, T_in, 32]
+
+                energy_contour_outputs *= energy_scales
+                energy_embeddings = tf.expand_dims(
+                    energy_contour_outputs, axis=2)  # [N, T_in, 1]
+                energy_embeddings = tf.layers.conv1d(
+                    energy_embeddings,
+                    filters=hp.encoder_projection_units,
+                    kernel_size=9,
+                    padding='same',
+                    name='energy_embeddings')  # [N, T_in, 32]
+
+            encoder_outputs_ = encoder_outputs + pitch_embeddings + energy_embeddings
+
+            # duration
+            dur_inputs = tf.concat([
+                encoder_outputs_, embedded_inputs_speaker,
+                embedded_inputs_emotion
+            ], 2)
+            if input_mask is not None:
+                dur_inputs = dur_inputs * tf.expand_dims(input_mask, -1)
+            with tf.variable_scope('Duration_Predictor'):
+                duration_predictor_cell = MultiRNNCell([
+                    LSTMBlockCell(hp.predictor_lstm_units),
+                    LSTMBlockCell(hp.predictor_lstm_units)
+                ], state_is_tuple=True)  # yapf:disable
+                duration_output_cell = DurPredictorCell(
+                    duration_predictor_cell, is_training, 1,
+                    hp.predictor_prenet_units)
+                duration_predictor_init_state = duration_output_cell.zero_state(
+                    batch_size=batch_size, dtype=tf.float32)
+                if is_training:
+                    duration_helper = VarTrainingHelper(
+                        tf.expand_dims(
+                            tf.log(tf.cast(durations, tf.float32) + 1),
+                            axis=2), dur_inputs, 1)
+                else:
+                    duration_helper = VarTestHelper(batch_size, dur_inputs, 1)
+                (
+                    duration_outputs, _
+                ), final_duration_predictor_state, _ = tf.contrib.seq2seq.dynamic_decode(
+                    BasicDecoder(duration_output_cell, duration_helper,
+                                 duration_predictor_init_state),
+                    maximum_iterations=1000)
+                duration_outputs = tf.squeeze(
+                    duration_outputs, axis=2)  # [N, T_in]
+                if input_mask is not None:
+                    duration_outputs = duration_outputs * input_mask
+                duration_outputs_ = tf.exp(duration_outputs) - 1
+
+            # Length Regulator
+            with tf.variable_scope('Length_Regulator'):
+                if is_training:
+                    i = tf.constant(1)
+                    # position embedding
+                    j = tf.constant(1)
+                    dur_len = tf.shape(durations)[-1]
+                    embedded_position_i = tf.range(1, durations[0, 0] + 1)
+
+                    def condition_pos(j, e):
+                        return tf.less(j, dur_len)
+
+                    def loop_body_pos(j, embedded_position_i):
+                        embedded_position_i = tf.concat([
+                            embedded_position_i,
+                            tf.range(1, durations[0, j] + 1)
+                        ], axis=0)  # yapf:disable
+                        return [j + 1, embedded_position_i]
+
+                    j, embedded_position_i = tf.while_loop(
+                        condition_pos,
+                        loop_body_pos, [j, embedded_position_i],
+                        shape_invariants=[
+                            j.get_shape(),
+                            tf.TensorShape([None])
+                        ])
+                    embedded_position = tf.reshape(embedded_position_i,
+                                                   (1, -1))
+
+                    # others
+                    LR_outputs = repeat(
+                        encoder_outputs_[0:1, :, :], durations[0, :], axis=1)
+                    embedded_outputs_speaker = repeat(
+                        embedded_inputs_speaker[0:1, :, :],
+                        durations[0, :],
+                        axis=1)
+                    embedded_outputs_emotion = repeat(
+                        embedded_inputs_emotion[0:1, :, :],
+                        durations[0, :],
+                        axis=1)
+
+                    def condition(i, pos, layer, s, e):
+                        return tf.less(i, tf.shape(mel_targets)[0])
+
+                    def loop_body(i, embedded_position, LR_outputs,
+                                  embedded_outputs_speaker,
+                                  embedded_outputs_emotion):
+                        # position embedding
+                        jj = tf.constant(1)
+                        embedded_position_i = tf.range(1, durations[i, 0] + 1)
+
+                        def condition_pos_i(j, e):
+                            return tf.less(j, dur_len)
+
+                        def loop_body_pos_i(j, embedded_position_i):
+                            embedded_position_i = tf.concat([
+                                embedded_position_i,
+                                tf.range(1, durations[i, j] + 1)
+                            ], axis=0)  # yapf:disable
+                            return [j + 1, embedded_position_i]
+
+                        jj, embedded_position_i = tf.while_loop(
+                            condition_pos_i,
+                            loop_body_pos_i, [jj, embedded_position_i],
+                            shape_invariants=[
+                                jj.get_shape(),
+                                tf.TensorShape([None])
+                            ])
+                        embedded_position = tf.concat([
+                            embedded_position,
+                            tf.reshape(embedded_position_i, (1, -1))
+                        ], 0)
+
+                        # others
+                        LR_outputs = tf.concat([
+                            LR_outputs,
+                            repeat(
+                                encoder_outputs_[i:i + 1, :, :],
+                                durations[i, :],
+                                axis=1)
+                        ], 0)
+                        embedded_outputs_speaker = tf.concat([
+                            embedded_outputs_speaker,
+                            repeat(
+                                embedded_inputs_speaker[i:i + 1, :, :],
+                                durations[i, :],
+                                axis=1)
+                        ], 0)
+                        embedded_outputs_emotion = tf.concat([
+                            embedded_outputs_emotion,
+                            repeat(
+                                embedded_inputs_emotion[i:i + 1, :, :],
+                                durations[i, :],
+                                axis=1)
+                        ], 0)
+                        return [
+                            i + 1, embedded_position, LR_outputs,
+                            embedded_outputs_speaker, embedded_outputs_emotion
+                        ]
+
+                    i, embedded_position, LR_outputs,
+                    embedded_outputs_speaker,
+                    embedded_outputs_emotion = tf.while_loop(
+                        condition,
+                        loop_body, [
+                            i, embedded_position, LR_outputs,
+                            embedded_outputs_speaker, embedded_outputs_emotion
+                        ],
+                        shape_invariants=[
+                            i.get_shape(),
+                            tf.TensorShape([None, None]),
+                            tf.TensorShape([None, None, None]),
+                            tf.TensorShape([None, None, None]),
+                            tf.TensorShape([None, None, None])
+                        ],
+                        parallel_iterations=hp.batch_size)
+
+                    ori_framenum = tf.shape(mel_targets)[1]
+                else:
+                    # position
+                    j = tf.constant(1)
+                    dur_len = tf.shape(duration_outputs_)[-1]
+                    embedded_position_i = tf.range(
+                        1,
+                        tf.cast(tf.round(duration_outputs_)[0, 0], tf.int32)
+                        + 1)
+
+                    def condition_pos(j, e):
+                        return tf.less(j, dur_len)
+
+                    def loop_body_pos(j, embedded_position_i):
+                        embedded_position_i = tf.concat([
+                            embedded_position_i,
+                            tf.range(
+                                1,
+                                tf.cast(
+                                    tf.round(duration_outputs_)[0, j],
+                                    tf.int32) + 1)
+                        ], axis=0)  # yapf:disable
+                        return [j + 1, embedded_position_i]
+
+                    j, embedded_position_i = tf.while_loop(
+                        condition_pos,
+                        loop_body_pos, [j, embedded_position_i],
+                        shape_invariants=[
+                            j.get_shape(),
+                            tf.TensorShape([None])
+                        ])
+                    embedded_position = tf.reshape(embedded_position_i,
+                                                   (1, -1))
+                    # others
+                    duration_outputs_ *= duration_scales
+                    LR_outputs = repeat(
+                        encoder_outputs_[0:1, :, :],
+                        tf.cast(tf.round(duration_outputs_)[0, :], tf.int32),
+                        axis=1)
+                    embedded_outputs_speaker = repeat(
+                        embedded_inputs_speaker[0:1, :, :],
+                        tf.cast(tf.round(duration_outputs_)[0, :], tf.int32),
+                        axis=1)
+                    embedded_outputs_emotion = repeat(
+                        embedded_inputs_emotion[0:1, :, :],
+                        tf.cast(tf.round(duration_outputs_)[0, :], tf.int32),
+                        axis=1)
+                    ori_framenum = tf.shape(LR_outputs)[1]
+
+                    left = hp.outputs_per_step - tf.mod(
+                        ori_framenum, hp.outputs_per_step)
+                    LR_outputs = tf.cond(
+                        tf.equal(left,
+                                 hp.outputs_per_step), lambda: LR_outputs,
+                        lambda: tf.pad(LR_outputs, [[0, 0], [0, left], [0, 0]],
+                                       'CONSTANT'))
+                    embedded_outputs_speaker = tf.cond(
+                        tf.equal(left, hp.outputs_per_step),
+                        lambda: embedded_outputs_speaker, lambda: tf.pad(
+                            embedded_outputs_speaker, [[0, 0], [0, left],
+                                                       [0, 0]], 'CONSTANT'))
+                    embedded_outputs_emotion = tf.cond(
+                        tf.equal(left, hp.outputs_per_step),
+                        lambda: embedded_outputs_emotion, lambda: tf.pad(
+                            embedded_outputs_emotion, [[0, 0], [0, left],
+                                                       [0, 0]], 'CONSTANT'))
+                    embedded_position = tf.cond(
+                        tf.equal(left, hp.outputs_per_step),
+                        lambda: embedded_position,
+                        lambda: tf.pad(embedded_position, [[0, 0], [0, left]],
+                                       'CONSTANT'))
+
+            # Pos_Embedding
+            with tf.variable_scope('Position_Embedding'):
+                Pos_Embedding = BatchSinusodalPositionalEncoding()
+                position_embeddings = Pos_Embedding.positional_encoding(
+                    batch_size,
+                    tf.shape(LR_outputs)[1], hp.encoder_projection_units,
+                    embedded_position)
+            LR_outputs += position_embeddings
+
+            # multi-frame
+            LR_outputs = tf.reshape(LR_outputs, [
+                batch_size, -1,
+                hp.outputs_per_step * hp.encoder_projection_units
+            ])
+            embedded_outputs_speaker = tf.reshape(
+                embedded_outputs_speaker,
+                [batch_size, -1, hp.outputs_per_step * 32])[:, :, :32]
+            embedded_outputs_emotion = tf.reshape(
+                embedded_outputs_emotion,
+                [batch_size, -1, hp.outputs_per_step * 32])[:, :, :32]
+            # [N, T_out, D_LR_outputs] (D_LR_outputs = hp.outputs_per_step * hp.encoder_projection_units + 64)
+            LR_outputs = tf.concat([
+                LR_outputs, embedded_outputs_speaker, embedded_outputs_emotion
+            ], -1)
+
+            # auto bandwidth
+            if is_training:
+                durations_mask = tf.cast(durations,
+                                         tf.float32) * input_mask  # [N, T_in]
+            else:
+                durations_mask = duration_outputs_
+            X_band_width = tf.cast(
+                tf.round(tf.reduce_max(durations_mask) / hp.outputs_per_step),
+                tf.int32)
+            H_band_width = X_band_width
+
+            with tf.variable_scope('Decoder'):
+                Decoder = SelfAttentionDecoder(
+                    num_layers=hp.decoder_num_layers,
+                    num_units=hp.decoder_num_units,
+                    num_heads=hp.decoder_num_heads,
+                    ffn_inner_dim=hp.decoder_ffn_inner_dim,
+                    dropout=hp.decoder_dropout,
+                    attention_dropout=hp.decoder_attention_dropout,
+                    relu_dropout=hp.decoder_relu_dropout,
+                    prenet_units=hp.prenet_units,
+                    dense_units=hp.prenet_proj_units,
+                    num_mels=hp.num_mels,
+                    outputs_per_step=hp.outputs_per_step,
+                    X_band_width=X_band_width,
+                    H_band_width=H_band_width,
+                    position_encoder=None)
+                if is_training:
+                    if hp.free_run:
+                        r = hp.outputs_per_step
+                        init_decoder_input = tf.expand_dims(
+                            tf.tile([[0.0]], [batch_size, hp.num_mels]),
+                            axis=1)  # [N, 1, hp.num_mels]
+                        decoder_input_lengths = tf.cast(
+                            output_lengths / r, tf.int32)
+                        decoder_outputs, attention_x, attention_h = Decoder.dynamic_decode_and_search(
+                            init_decoder_input,
+                            maximum_iterations=tf.shape(LR_outputs)[1],
+                            mode=is_training,
+                            memory=LR_outputs,
+                            memory_sequence_length=decoder_input_lengths)
+                    else:
+                        r = hp.outputs_per_step
+                        decoder_input = mel_targets[:, r - 1::
+                                                    r, :]  # [N, T_out / r, hp.num_mels]
+                        init_decoder_input = tf.expand_dims(
+                            tf.tile([[0.0]], [batch_size, hp.num_mels]),
+                            axis=1)  # [N, 1, hp.num_mels]
+                        decoder_input = tf.concat(
+                            [init_decoder_input, decoder_input],
+                            axis=1)  # [N, T_out / r + 1, hp.num_mels]
+                        decoder_input = decoder_input[:, :
+                                                      -1, :]  # [N, T_out / r, hp.num_mels]
+                        decoder_input_lengths = tf.cast(
+                            output_lengths / r, tf.int32)
+                        decoder_outputs, attention_x, attention_h = Decoder.decode_from_inputs(
+                            decoder_input,
+                            decoder_input_lengths,
+                            mode=is_training,
+                            memory=LR_outputs,
+                            memory_sequence_length=decoder_input_lengths)
+                else:
+                    init_decoder_input = tf.expand_dims(
+                        tf.tile([[0.0]], [batch_size, hp.num_mels]),
+                        axis=1)  # [N, 1, hp.num_mels]
+                    decoder_outputs, attention_x, attention_h = Decoder.dynamic_decode_and_search(
+                        init_decoder_input,
+                        maximum_iterations=tf.shape(LR_outputs)[1],
+                        mode=is_training,
+                        memory=LR_outputs,
+                        memory_sequence_length=tf.expand_dims(
+                            tf.shape(LR_outputs)[1], axis=0))
+
+                if is_training:
+                    mel_outputs_ = tf.reshape(decoder_outputs,
+                                              [batch_size, -1, hp.num_mels])
+                else:
+                    mel_outputs_ = tf.reshape(
+                        decoder_outputs,
+                        [batch_size, -1, hp.num_mels])[:, :ori_framenum, :]
+                mel_outputs = mel_outputs_
+
+            with tf.variable_scope('Postnet'):
+                Postnet_FSMN = FsmnEncoderV2(
+                    filter_size=hp.postnet_filter_size,
+                    fsmn_num_layers=hp.postnet_fsmn_num_layers,
+                    dnn_num_layers=hp.postnet_dnn_num_layers,
+                    num_memory_units=hp.postnet_num_memory_units,
+                    ffn_inner_dim=hp.postnet_ffn_inner_dim,
+                    dropout=hp.postnet_dropout,
+                    shift=hp.postnet_shift,
+                    position_encoder=None)
+                if is_training:
+                    postnet_fsmn_outputs, _, _ = Postnet_FSMN.encode(
+                        mel_outputs,
+                        sequence_length=output_lengths,
+                        mode=is_training)
+                    hidden_lstm_outputs, _ = tf.nn.dynamic_rnn(
+                        LSTMBlockCell(hp.postnet_lstm_units),
+                        postnet_fsmn_outputs,
+                        sequence_length=output_lengths,
+                        dtype=tf.float32)
+                else:
+                    postnet_fsmn_outputs, _, _ = Postnet_FSMN.encode(
+                        mel_outputs,
+                        sequence_length=[tf.shape(mel_outputs_)[1]],
+                        mode=is_training)
+                    hidden_lstm_outputs, _ = tf.nn.dynamic_rnn(
+                        LSTMBlockCell(hp.postnet_lstm_units),
+                        postnet_fsmn_outputs,
+                        sequence_length=[tf.shape(mel_outputs_)[1]],
+                        dtype=tf.float32)
+
+            mel_residual_outputs = tf.layers.dense(
+                hidden_lstm_outputs, units=hp.num_mels)
+            mel_outputs += mel_residual_outputs
+
+            self.inputs = inputs
+            self.inputs_speaker = inputs_speaker
+            self.inputs_emotion = inputs_emotion
+            self.input_lengths = input_lengths
+            self.durations = durations
+            self.output_lengths = output_lengths
+            self.mel_outputs_ = mel_outputs_
+            self.mel_outputs = mel_outputs
+            self.mel_targets = mel_targets
+            self.duration_outputs = duration_outputs
+            self.duration_outputs_ = duration_outputs_
+            self.duration_scales = duration_scales
+            self.pitch_contour_outputs = pitch_contour_outputs
+            self.pitch_contours = pitch_contours
+            self.pitch_scales = pitch_scales
+            self.energy_contour_outputs = energy_contour_outputs
+            self.energy_contours = energy_contours
+            self.energy_scales = energy_scales
+            self.uv_masks_ = uv_masks
+
+            self.embedded_inputs_emotion = embedded_inputs_emotion
+            self.embedding_fsmn_outputs = embedded_inputs
+            self.encoder_outputs = encoder_outputs
+            self.encoder_outputs_ = encoder_outputs_
+            self.LR_outputs = LR_outputs
+            self.postnet_fsmn_outputs = postnet_fsmn_outputs
+
+            self.pitch_embeddings = pitch_embeddings
+            self.energy_embeddings = energy_embeddings
+
+            self.attns = attns
+            self.attention_x = attention_x
+            self.attention_h = attention_h
+            self.X_band_width = X_band_width
+            self.H_band_width = H_band_width
+
+    def add_loss(self):
+        '''Adds loss to the model. Sets "loss" field. initialize must have been called.'''
+        with tf.variable_scope('loss') as _:
+            hp = self._hparams
+            mask = tf.sequence_mask(
+                self.output_lengths,
+                tf.shape(self.mel_targets)[1],
+                dtype=tf.float32)
+            valid_outputs = tf.reduce_sum(mask)
+
+            mask_input = tf.sequence_mask(
+                self.input_lengths,
+                tf.shape(self.durations)[1],
+                dtype=tf.float32)
+            valid_inputs = tf.reduce_sum(mask_input)
+
+            # mel loss
+            if self.uv_masks_ is not None:
+                valid_outputs_mask = tf.reduce_sum(
+                    tf.expand_dims(mask, -1) * self.uv_masks_)
+                self.mel_loss_ = tf.reduce_sum(
+                    tf.abs(self.mel_targets - self.mel_outputs_)
+                    * tf.expand_dims(mask, -1) * self.uv_masks_) / (
+                        valid_outputs_mask * hp.num_mels)
+                self.mel_loss = tf.reduce_sum(
+                    tf.abs(self.mel_targets - self.mel_outputs)
+                    * tf.expand_dims(mask, -1) * self.uv_masks_) / (
+                        valid_outputs_mask * hp.num_mels)
+            else:
+                self.mel_loss_ = tf.reduce_sum(
+                    tf.abs(self.mel_targets - self.mel_outputs_)
+                    * tf.expand_dims(mask, -1)) / (
+                        valid_outputs * hp.num_mels)
+                self.mel_loss = tf.reduce_sum(
+                    tf.abs(self.mel_targets - self.mel_outputs)
+                    * tf.expand_dims(mask, -1)) / (
+                        valid_outputs * hp.num_mels)
+
+            # duration loss
+            self.duration_loss = tf.reduce_sum(
+                tf.abs(
+                    tf.log(tf.cast(self.durations, tf.float32) + 1)
+                    - self.duration_outputs) * mask_input) / valid_inputs
+
+            # pitch contour loss
+            self.pitch_contour_loss = tf.reduce_sum(
+                tf.abs(self.pitch_contours - self.pitch_contour_outputs)
+                * mask_input) / valid_inputs
+
+            # energy contour loss
+            self.energy_contour_loss = tf.reduce_sum(
+                tf.abs(self.energy_contours - self.energy_contour_outputs)
+                * mask_input) / valid_inputs
+
+            # final loss
+            self.loss = self.mel_loss_ + self.mel_loss + self.duration_loss \
+                + self.pitch_contour_loss + self.energy_contour_loss
+
+            # guided attention loss
+            self.guided_attention_loss = tf.constant(0.0)
+            if hp.guided_attention:
+                i0 = tf.constant(0)
+                loss0 = tf.constant(0.0)
+
+                def c(i, _):
+                    return tf.less(i, tf.shape(mel_targets)[0])
+
+                def loop_body(i, loss):
+                    decoder_input_lengths = tf.cast(
+                        self.output_lengths / hp.outputs_per_step, tf.int32)
+                    input_len = decoder_input_lengths[i]
+                    output_len = decoder_input_lengths[i]
+                    input_w = tf.expand_dims(
+                        tf.range(tf.cast(input_len, dtype=tf.float32)),
+                        axis=1) / tf.cast(
+                            input_len, dtype=tf.float32)  # [T_in, 1]
+                    output_w = tf.expand_dims(
+                        tf.range(tf.cast(output_len, dtype=tf.float32)),
+                        axis=0) / tf.cast(
+                            output_len, dtype=tf.float32)  # [1, T_out]
+                    guided_attention_w = 1.0 - tf.exp(
+                        -(1 / hp.guided_attention_2g_squared)
+                        * tf.square(input_w - output_w))  # [T_in, T_out]
+                    guided_attention_w = tf.expand_dims(
+                        guided_attention_w, axis=0)  # [1, T_in, T_out]
+                    # [hp.decoder_num_heads, T_in, T_out]
+                    guided_attention_w = tf.tile(guided_attention_w,
+                                                 [hp.decoder_num_heads, 1, 1])
+                    loss_i = tf.constant(0.0)
+                    for j in range(hp.decoder_num_layers):
+                        loss_i += tf.reduce_mean(
+                            self.attention_h[j][i, :, :input_len, :output_len]
+                            * guided_attention_w)
+
+                    return [tf.add(i, 1), tf.add(loss, loss_i)]
+
+                _, loss = tf.while_loop(
+                    c,
+                    loop_body,
+                    loop_vars=[i0, loss0],
+                    parallel_iterations=hp.batch_size)
+                self.guided_attention_loss = loss / hp.batch_size
+                self.loss += hp.guided_attention_loss_weight * self.guided_attention_loss
+
+    def add_optimizer(self, global_step):
+        '''Adds optimizer. Sets "gradients" and "optimize" fields. add_loss must have been called.
+
+        Args:
+          global_step: int32 scalar Tensor representing current global step in training
+        '''
+        with tf.variable_scope('optimizer') as _:
+            hp = self._hparams
+            if hp.decay_learning_rate:
+                self.learning_rate = _learning_rate_decay(
+                    hp.initial_learning_rate, global_step)
+            else:
+                self.learning_rate = tf.convert_to_tensor(
+                    hp.initial_learning_rate)
+            optimizer = tf.train.AdamOptimizer(self.learning_rate,
+                                               hp.adam_beta1, hp.adam_beta2)
+            gradients, variables = zip(*optimizer.compute_gradients(self.loss))
+            self.gradients = gradients
+            clipped_gradients, _ = tf.clip_by_global_norm(gradients, 1.0)
+
+            # Add dependency on UPDATE_OPS; otherwise batchnorm won't work correctly. See:
+            # https://github.com/tensorflow/tensorflow/issues/1122
+            with tf.control_dependencies(
+                    tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
+                self.optimize = optimizer.apply_gradients(
+                    zip(clipped_gradients, variables), global_step=global_step)
+
+
+def _learning_rate_decay(init_lr, global_step):
+    # Noam scheme from tensor2tensor:
+    warmup_steps = 4000.0
+    step = tf.cast(global_step + 1, dtype=tf.float32)
+    return init_lr * warmup_steps**0.5 * tf.minimum(step * warmup_steps**-1.5,
+                                                    step**-0.5)

modelscope/models/audio/tts/am/models/self_attention_decoder.py

@@ -0,0 +1,817 @@
+"""Define self-attention decoder."""
+
+import sys
+
+import tensorflow as tf
+
+from . import compat, transformer
+from .modules import decoder_prenet
+from .position import SinusoidalPositionEncoder
+
+
+class SelfAttentionDecoder():
+    """Decoder using self-attention as described in
+    https://arxiv.org/abs/1706.03762.
+    """
+
+    def __init__(self,
+                 num_layers,
+                 num_units=512,
+                 num_heads=8,
+                 ffn_inner_dim=2048,
+                 dropout=0.1,
+                 attention_dropout=0.1,
+                 relu_dropout=0.1,
+                 prenet_units=256,
+                 dense_units=128,
+                 num_mels=80,
+                 outputs_per_step=3,
+                 X_band_width=None,
+                 H_band_width=None,
+                 position_encoder=SinusoidalPositionEncoder(),
+                 self_attention_type='scaled_dot'):
+        """Initializes the parameters of the decoder.
+
+        Args:
+          num_layers: The number of layers.
+          num_units: The number of hidden units.
+          num_heads: The number of heads in the multi-head attention.
+          ffn_inner_dim: The number of units of the inner linear transformation
+            in the feed forward layer.
+          dropout: The probability to drop units from the outputs.
+          attention_dropout: The probability to drop units from the attention.
+          relu_dropout: The probability to drop units from the ReLU activation in
+            the feed forward layer.
+          position_encoder: A :class:`opennmt.layers.position.PositionEncoder` to
+            apply on inputs or ``None``.
+          self_attention_type: Type of self attention, "scaled_dot" or "average" (case
+            insensitive).
+
+        Raises:
+          ValueError: if :obj:`self_attention_type` is invalid.
+        """
+        super(SelfAttentionDecoder, self).__init__()
+        self.num_layers = num_layers
+        self.num_units = num_units
+        self.num_heads = num_heads
+        self.ffn_inner_dim = ffn_inner_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.relu_dropout = relu_dropout
+        self.position_encoder = position_encoder
+        self.self_attention_type = self_attention_type.lower()
+        if self.self_attention_type not in ('scaled_dot', 'average'):
+            raise ValueError('invalid attention type %s'
+                             % self.self_attention_type)
+        if self.self_attention_type == 'average':
+            tf.logging.warning(
+                'Support for average attention network is experimental '
+                'and may change in future versions.')
+        self.prenet_units = prenet_units
+        self.dense_units = dense_units
+        self.num_mels = num_mels
+        self.outputs_per_step = outputs_per_step
+        self.X_band_width = X_band_width
+        self.H_band_width = H_band_width
+
+    @property
+    def output_size(self):
+        """Returns the decoder output size."""
+        return self.num_units
+
+    @property
+    def support_alignment_history(self):
+        return True
+
+    @property
+    def support_multi_source(self):
+        return True
+
+    def _init_cache(self, batch_size, dtype=tf.float32, num_sources=1):
+        cache = {}
+
+        for layer in range(self.num_layers):
+            proj_cache_shape = [
+                batch_size, self.num_heads, 0, self.num_units // self.num_heads
+            ]
+            layer_cache = {}
+            layer_cache['memory'] = [{
+                'memory_keys':
+                tf.zeros(proj_cache_shape, dtype=dtype),
+                'memory_values':
+                tf.zeros(proj_cache_shape, dtype=dtype)
+            } for _ in range(num_sources)]
+            if self.self_attention_type == 'scaled_dot':
+                layer_cache['self_keys'] = tf.zeros(
+                    proj_cache_shape, dtype=dtype)
+                layer_cache['self_values'] = tf.zeros(
+                    proj_cache_shape, dtype=dtype)
+            elif self.self_attention_type == 'average':
+                layer_cache['prev_g'] = tf.zeros(
+                    [batch_size, 1, self.num_units], dtype=dtype)
+            cache['layer_{}'.format(layer)] = layer_cache
+
+        return cache
+
+    def _init_attn(self, dtype=tf.float32):
+        attn = []
+        for layer in range(self.num_layers):
+            attn.append(tf.TensorArray(tf.float32, size=0, dynamic_size=True))
+        return attn
+
+    def _self_attention_stack(self,
+                              inputs,
+                              sequence_length=None,
+                              mode=True,
+                              cache=None,
+                              memory=None,
+                              memory_sequence_length=None,
+                              step=None):
+
+        # [N, T_out, self.dense_units] or [N, 1, self.dense_units]
+        prenet_outputs = decoder_prenet(inputs, self.prenet_units,
+                                        self.dense_units, mode)
+        if step is None:
+            decoder_inputs = tf.concat(
+                [memory, prenet_outputs],
+                axis=-1)  # [N, T_out, memory_size + self.dense_units]
+        else:
+            decoder_inputs = tf.concat(
+                [memory[:, step:step + 1, :], prenet_outputs],
+                axis=-1)  # [N, 1, memory_size + self.dense_units]
+        decoder_inputs = tf.layers.dense(
+            decoder_inputs, units=self.dense_units)
+
+        inputs = decoder_inputs
+        inputs *= self.num_units**0.5
+        if self.position_encoder is not None:
+            inputs = self.position_encoder(
+                inputs, position=step + 1 if step is not None else None)
+
+        inputs = tf.layers.dropout(inputs, rate=self.dropout, training=mode)
+
+        decoder_mask = None
+        memory_mask = None
+        # last_attention = None
+
+        X_band_width_tmp = -1
+        H_band_width_tmp = -1
+        if self.X_band_width is not None:
+            X_band_width_tmp = tf.cast(
+                tf.cond(
+                    tf.less(tf.shape(memory)[1], self.X_band_width),
+                    lambda: -1, lambda: self.X_band_width),
+                dtype=tf.int64)
+        if self.H_band_width is not None:
+            H_band_width_tmp = tf.cast(
+                tf.cond(
+                    tf.less(tf.shape(memory)[1], self.H_band_width),
+                    lambda: -1, lambda: self.H_band_width),
+                dtype=tf.int64)
+
+        if self.self_attention_type == 'scaled_dot':
+            if sequence_length is not None:
+                decoder_mask = transformer.build_future_mask(
+                    sequence_length,
+                    num_heads=self.num_heads,
+                    maximum_length=tf.shape(inputs)[1],
+                    band=X_band_width_tmp)  # [N, 1, T_out, T_out]
+        elif self.self_attention_type == 'average':
+            if cache is None:
+                if sequence_length is None:
+                    sequence_length = tf.fill([tf.shape(inputs)[0]],
+                                              tf.shape(inputs)[1])
+                decoder_mask = transformer.cumulative_average_mask(
+                    sequence_length,
+                    maximum_length=tf.shape(inputs)[1],
+                    dtype=inputs.dtype)
+
+        if memory is not None and not tf.contrib.framework.nest.is_sequence(
+                memory):
+            memory = (memory, )
+        if memory_sequence_length is not None:
+            if not tf.contrib.framework.nest.is_sequence(
+                    memory_sequence_length):
+                memory_sequence_length = (memory_sequence_length, )
+            if step is None:
+                memory_mask = [
+                    transformer.build_history_mask(
+                        length,
+                        num_heads=self.num_heads,
+                        maximum_length=tf.shape(m)[1],
+                        band=H_band_width_tmp)
+                    for m, length in zip(memory, memory_sequence_length)
+                ]
+            else:
+                memory_mask = [
+                    transformer.build_history_mask(
+                        length,
+                        num_heads=self.num_heads,
+                        maximum_length=tf.shape(m)[1],
+                        band=H_band_width_tmp)[:, :, step:step + 1, :]
+                    for m, length in zip(memory, memory_sequence_length)
+                ]
+
+        # last_attention = None
+        attns_x = []
+        attns_h = []
+        for layer in range(self.num_layers):
+            layer_name = 'layer_{}'.format(layer)
+            layer_cache = cache[layer_name] if cache is not None else None
+            with tf.variable_scope(layer_name):
+                if memory is not None:
+                    for i, (mem, mask) in enumerate(zip(memory, memory_mask)):
+                        memory_cache = None
+                        if layer_cache is not None:
+                            memory_cache = layer_cache['memory'][i]
+                        scope_name = 'multi_head_{}'.format(i)
+                        if i == 0:
+                            scope_name = 'multi_head'
+                        with tf.variable_scope(scope_name):
+                            encoded, attn_x, attn_h = transformer.multi_head_attention_PNCA(
+                                self.num_heads,
+                                transformer.norm(inputs),
+                                mem,
+                                mode,
+                                num_units=self.num_units,
+                                mask=decoder_mask,
+                                mask_h=mask,
+                                cache=layer_cache,
+                                cache_h=memory_cache,
+                                dropout=self.attention_dropout,
+                                return_attention=True,
+                                layer_name=layer_name,
+                                X_band_width=self.X_band_width)
+                            attns_x.append(attn_x)
+                            attns_h.append(attn_h)
+                            context = transformer.drop_and_add(
+                                inputs, encoded, mode, dropout=self.dropout)
+
+                with tf.variable_scope('ffn'):
+                    transformed = transformer.feed_forward_ori(
+                        transformer.norm(context),
+                        self.ffn_inner_dim,
+                        mode,
+                        dropout=self.relu_dropout)
+                    transformed = transformer.drop_and_add(
+                        context, transformed, mode, dropout=self.dropout)
+
+                inputs = transformed
+
+        outputs = transformer.norm(inputs)
+        outputs = tf.layers.dense(
+            outputs, units=self.num_mels * self.outputs_per_step)
+        return outputs, attns_x, attns_h
+
+    def decode_from_inputs(self,
+                           inputs,
+                           sequence_length,
+                           initial_state=None,
+                           mode=True,
+                           memory=None,
+                           memory_sequence_length=None):
+        outputs, attention_x, attention_h = self._self_attention_stack(
+            inputs,
+            sequence_length=sequence_length,
+            mode=mode,
+            memory=memory,
+            memory_sequence_length=memory_sequence_length)
+        return outputs, attention_x, attention_h
+
+    def step_fn(self,
+                mode,
+                batch_size,
+                initial_state=None,
+                memory=None,
+                memory_sequence_length=None,
+                dtype=tf.float32):
+        if memory is None:
+            num_sources = 0
+        elif tf.contrib.framework.nest.is_sequence(memory):
+            num_sources = len(memory)
+        else:
+            num_sources = 1
+        cache = self._init_cache(
+            batch_size, dtype=dtype, num_sources=num_sources)
+        attention_x = self._init_attn(dtype=dtype)
+        attention_h = self._init_attn(dtype=dtype)
+
+        def _fn(step, inputs, cache):
+            outputs, attention_x, attention_h = self._self_attention_stack(
+                inputs,
+                mode=mode,
+                cache=cache,
+                memory=memory,
+                memory_sequence_length=memory_sequence_length,
+                step=step)
+            attention_x_tmp = []
+            for layer in range(len(attention_h)):
+                attention_x_tmp_l = tf.zeros_like(attention_h[layer])
+                if self.X_band_width is not None:
+                    pred = tf.less(step, self.X_band_width + 1)
+                    attention_x_tmp_l_1 = tf.cond(pred,  # yapf:disable
+                                                  lambda: attention_x_tmp_l[:, :, :, :step + 1] + attention_x[layer],
+                                                  lambda: tf.concat([
+                                                                    attention_x_tmp_l[:, :, :,
+                                                                                      :step - self.X_band_width],
+                                                                    attention_x_tmp_l[:, :, :,
+                                                                                      step - self.X_band_width:step + 1]
+                                                                    + attention_x[layer]],
+                                                                    axis=-1))  # yapf:disable
+                    attention_x_tmp_l_2 = attention_x_tmp_l[:, :, :, step + 1:]
+                    attention_x_tmp.append(
+                        tf.concat([attention_x_tmp_l_1, attention_x_tmp_l_2],
+                                  axis=-1))
+                else:
+                    attention_x_tmp_l_1 = attention_x_tmp_l[:, :, :, :step + 1]
+                    attention_x_tmp_l_2 = attention_x_tmp_l[:, :, :, step + 1:]
+                    attention_x_tmp.append(
+                        tf.concat([
+                            attention_x_tmp_l_1 + attention_x[layer],
+                            attention_x_tmp_l_2
+                        ], axis=-1))  # yapf:disable
+            attention_x = attention_x_tmp
+            return outputs, cache, attention_x, attention_h
+
+        return _fn, cache, attention_x, attention_h
+
+    def dynamic_decode_and_search(self, init_decoder_input, maximum_iterations,
+                                  mode, memory, memory_sequence_length):
+        batch_size = tf.shape(init_decoder_input)[0]
+        step_fn, init_cache, init_attn_x, init_attn_h = self.step_fn(
+            mode,
+            batch_size,
+            memory=memory,
+            memory_sequence_length=memory_sequence_length)
+
+        outputs, attention_x, attention_h, cache = self.dynamic_decode(
+            step_fn,
+            init_decoder_input,
+            init_cache=init_cache,
+            init_attn_x=init_attn_x,
+            init_attn_h=init_attn_h,
+            maximum_iterations=maximum_iterations,
+            batch_size=batch_size)
+        return outputs, attention_x, attention_h
+
+    def dynamic_decode_and_search_teacher_forcing(self, decoder_input,
+                                                  maximum_iterations, mode,
+                                                  memory,
+                                                  memory_sequence_length):
+        batch_size = tf.shape(decoder_input)[0]
+        step_fn, init_cache, init_attn_x, init_attn_h = self.step_fn(
+            mode,
+            batch_size,
+            memory=memory,
+            memory_sequence_length=memory_sequence_length)
+
+        outputs, attention_x, attention_h, cache = self.dynamic_decode_teacher_forcing(
+            step_fn,
+            decoder_input,
+            init_cache=init_cache,
+            init_attn_x=init_attn_x,
+            init_attn_h=init_attn_h,
+            maximum_iterations=maximum_iterations,
+            batch_size=batch_size)
+        return outputs, attention_x, attention_h
+
+    def dynamic_decode(self,
+                       step_fn,
+                       init_decoder_input,
+                       init_cache=None,
+                       init_attn_x=None,
+                       init_attn_h=None,
+                       maximum_iterations=None,
+                       batch_size=None):
+
+        def _cond(step, cache, inputs, outputs, attention_x, attention_h):  # pylint: disable=unused-argument
+            return tf.less(step, maximum_iterations)
+
+        def _body(step, cache, inputs, outputs, attention_x, attention_h):
+            # output: [1, 1, num_mels * r]
+            # attn: [1, 1, T_out]
+            output, cache, attn_x, attn_h = step_fn(
+                step, inputs, cache)  # outputs, cache, attention, attns
+            for layer in range(len(attention_x)):
+                attention_x[layer] = attention_x[layer].write(
+                    step, tf.cast(attn_x[layer], tf.float32))
+
+            for layer in range(len(attention_h)):
+                attention_h[layer] = attention_h[layer].write(
+                    step, tf.cast(attn_h[layer], tf.float32))
+
+            outputs = outputs.write(step, tf.cast(output, tf.float32))
+            return step + 1, cache, output[:, :, -self.
+                                           num_mels:], outputs, attention_x, attention_h
+
+        step = tf.constant(0, dtype=tf.int32)
+        outputs = tf.TensorArray(tf.float32, size=0, dynamic_size=True)
+
+        _, cache, _, outputs, attention_x, attention_h = tf.while_loop(
+            _cond,
+            _body,
+            loop_vars=(step, init_cache, init_decoder_input, outputs,
+                       init_attn_x, init_attn_h),
+            shape_invariants=(step.shape,
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants, init_cache),
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants,
+                                  init_decoder_input), tf.TensorShape(None),
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants, init_attn_x),
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants, init_attn_h)),
+            parallel_iterations=1,
+            back_prop=False,
+            maximum_iterations=maximum_iterations)
+        # element of outputs: [N, 1, num_mels * r]
+        outputs_stack = outputs.stack()  # [T_out, N, 1, num_mels * r]
+        outputs_stack = tf.transpose(
+            outputs_stack, perm=[2, 1, 0, 3])  # [1, N, T_out, num_mels * r]
+        outputs_stack = tf.squeeze(
+            outputs_stack, axis=0)  # [N, T_out, num_mels * r]
+
+        attention_x_stack = []
+        for layer in range(len(attention_x)):
+            attention_x_stack_tmp = attention_x[layer].stack(
+            )  # [T_out, N, H, 1, T_out]
+            attention_x_stack_tmp = tf.transpose(
+                attention_x_stack_tmp, perm=[3, 1, 2, 0,
+                                             4])  # [1, N, H, T_out, T_out]
+            attention_x_stack_tmp = tf.squeeze(
+                attention_x_stack_tmp, axis=0)  # [N, H, T_out, T_out]
+            attention_x_stack.append(attention_x_stack_tmp)
+
+        attention_h_stack = []
+        for layer in range(len(attention_h)):
+            attention_h_stack_tmp = attention_h[layer].stack(
+            )  # [T_out, N, H, 1, T_out]
+            attention_h_stack_tmp = tf.transpose(
+                attention_h_stack_tmp, perm=[3, 1, 2, 0,
+                                             4])  # [1, N, H, T_out, T_out]
+            attention_h_stack_tmp = tf.squeeze(
+                attention_h_stack_tmp, axis=0)  # [N, H, T_out, T_out]
+            attention_h_stack.append(attention_h_stack_tmp)
+
+        return outputs_stack, attention_x_stack, attention_h_stack, cache
+
+    def dynamic_decode_teacher_forcing(self,
+                                       step_fn,
+                                       decoder_input,
+                                       init_cache=None,
+                                       init_attn_x=None,
+                                       init_attn_h=None,
+                                       maximum_iterations=None,
+                                       batch_size=None):
+
+        def _cond(step, cache, inputs, outputs, attention_x, attention_h):  # pylint: disable=unused-argument
+            return tf.less(step, maximum_iterations)
+
+        def _body(step, cache, inputs, outputs, attention_x, attention_h):
+            # output: [1, 1, num_mels * r]
+            # attn: [1, 1, T_out]
+            output, cache, attn_x, attn_h = step_fn(
+                step, inputs[:, step:step + 1, :],
+                cache)  # outputs, cache, attention, attns
+            for layer in range(len(attention_x)):
+                attention_x[layer] = attention_x[layer].write(
+                    step, tf.cast(attn_x[layer], tf.float32))
+
+            for layer in range(len(attention_h)):
+                attention_h[layer] = attention_h[layer].write(
+                    step, tf.cast(attn_h[layer], tf.float32))
+            outputs = outputs.write(step, tf.cast(output, tf.float32))
+            return step + 1, cache, inputs, outputs, attention_x, attention_h
+
+        step = tf.constant(0, dtype=tf.int32)
+        outputs = tf.TensorArray(tf.float32, size=0, dynamic_size=True)
+
+        _, cache, _, outputs, attention_x, attention_h = tf.while_loop(
+            _cond,
+            _body,
+            loop_vars=(step, init_cache, decoder_input, outputs, init_attn_x,
+                       init_attn_h),
+            shape_invariants=(step.shape,
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants,
+                                  init_cache), decoder_input.shape,
+                              tf.TensorShape(None),
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants, init_attn_x),
+                              compat.nest.map_structure(
+                                  self._get_shape_invariants, init_attn_h)),
+            parallel_iterations=1,
+            back_prop=False,
+            maximum_iterations=maximum_iterations)
+        # element of outputs: [N, 1, num_mels * r]
+        outputs_stack = outputs.stack()  # [T_out, N, 1, num_mels * r]
+        outputs_stack = tf.transpose(
+            outputs_stack, perm=[2, 1, 0, 3])  # [1, N, T_out, num_mels * r]
+        outputs_stack = tf.squeeze(
+            outputs_stack, axis=0)  # [N, T_out, num_mels * r]
+
+        attention_x_stack = []
+        for layer in range(len(attention_x)):
+            attention_x_stack_tmp = attention_x[layer].stack(
+            )  # [T_out, N, H, 1, T_out]
+            attention_x_stack_tmp = tf.transpose(
+                attention_x_stack_tmp, perm=[3, 1, 2, 0,
+                                             4])  # [1, N, H, T_out, T_out]
+            attention_x_stack_tmp = tf.squeeze(
+                attention_x_stack_tmp, axis=0)  # [N, H, T_out, T_out]
+            attention_x_stack.append(attention_x_stack_tmp)
+
+        attention_h_stack = []
+        for layer in range(len(attention_h)):
+            attention_h_stack_tmp = attention_h[layer].stack(
+            )  # [T_out, N, H, 1, T_out]
+            attention_h_stack_tmp = tf.transpose(
+                attention_h_stack_tmp, perm=[3, 1, 2, 0,
+                                             4])  # [1, N, H, T_out, T_out]
+            attention_h_stack_tmp = tf.squeeze(
+                attention_h_stack_tmp, axis=0)  # [N, H, T_out, T_out]
+            attention_h_stack.append(attention_h_stack_tmp)
+
+        return outputs_stack, attention_x_stack, attention_h_stack, cache
+
+    def _get_shape_invariants(self, tensor):
+        """Returns the shape of the tensor but sets middle dims to None."""
+        if isinstance(tensor, tf.TensorArray):
+            shape = None
+        else:
+            shape = tensor.shape.as_list()
+            for i in range(1, len(shape) - 1):
+                shape[i] = None
+        return tf.TensorShape(shape)
+
+
+class SelfAttentionDecoderOri():
+    """Decoder using self-attention as described in
+    https://arxiv.org/abs/1706.03762.
+    """
+
+    def __init__(self,
+                 num_layers,
+                 num_units=512,
+                 num_heads=8,
+                 ffn_inner_dim=2048,
+                 dropout=0.1,
+                 attention_dropout=0.1,
+                 relu_dropout=0.1,
+                 position_encoder=SinusoidalPositionEncoder(),
+                 self_attention_type='scaled_dot'):
+        """Initializes the parameters of the decoder.
+
+        Args:
+          num_layers: The number of layers.
+          num_units: The number of hidden units.
+          num_heads: The number of heads in the multi-head attention.
+          ffn_inner_dim: The number of units of the inner linear transformation
+            in the feed forward layer.
+          dropout: The probability to drop units from the outputs.
+          attention_dropout: The probability to drop units from the attention.
+          relu_dropout: The probability to drop units from the ReLU activation in
+            the feed forward layer.
+          position_encoder: A :class:`opennmt.layers.position.PositionEncoder` to
+            apply on inputs or ``None``.
+          self_attention_type: Type of self attention, "scaled_dot" or "average" (case
+            insensitive).
+
+        Raises:
+          ValueError: if :obj:`self_attention_type` is invalid.
+        """
+        super(SelfAttentionDecoderOri, self).__init__()
+        self.num_layers = num_layers
+        self.num_units = num_units
+        self.num_heads = num_heads
+        self.ffn_inner_dim = ffn_inner_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.relu_dropout = relu_dropout
+        self.position_encoder = position_encoder
+        self.self_attention_type = self_attention_type.lower()
+        if self.self_attention_type not in ('scaled_dot', 'average'):
+            raise ValueError('invalid attention type %s'
+                             % self.self_attention_type)
+        if self.self_attention_type == 'average':
+            tf.logging.warning(
+                'Support for average attention network is experimental '
+                'and may change in future versions.')
+
+    @property
+    def output_size(self):
+        """Returns the decoder output size."""
+        return self.num_units
+
+    @property
+    def support_alignment_history(self):
+        return True
+
+    @property
+    def support_multi_source(self):
+        return True
+
+    def _init_cache(self, batch_size, dtype=tf.float32, num_sources=1):
+        cache = {}
+
+        for layer in range(self.num_layers):
+            proj_cache_shape = [
+                batch_size, self.num_heads, 0, self.num_units // self.num_heads
+            ]
+            layer_cache = {}
+            layer_cache['memory'] = [{
+                'memory_keys':
+                tf.zeros(proj_cache_shape, dtype=dtype),
+                'memory_values':
+                tf.zeros(proj_cache_shape, dtype=dtype)
+            } for _ in range(num_sources)]
+            if self.self_attention_type == 'scaled_dot':
+                layer_cache['self_keys'] = tf.zeros(
+                    proj_cache_shape, dtype=dtype)
+                layer_cache['self_values'] = tf.zeros(
+                    proj_cache_shape, dtype=dtype)
+            elif self.self_attention_type == 'average':
+                layer_cache['prev_g'] = tf.zeros(
+                    [batch_size, 1, self.num_units], dtype=dtype)
+            cache['layer_{}'.format(layer)] = layer_cache
+
+        return cache
+
+    def _self_attention_stack(self,
+                              inputs,
+                              sequence_length=None,
+                              mode=True,
+                              cache=None,
+                              memory=None,
+                              memory_sequence_length=None,
+                              step=None):
+        inputs *= self.num_units**0.5
+        if self.position_encoder is not None:
+            inputs = self.position_encoder(
+                inputs, position=step + 1 if step is not None else None)
+
+        inputs = tf.layers.dropout(inputs, rate=self.dropout, training=mode)
+
+        decoder_mask = None
+        memory_mask = None
+        last_attention = None
+
+        if self.self_attention_type == 'scaled_dot':
+            if sequence_length is not None:
+                decoder_mask = transformer.build_future_mask(
+                    sequence_length,
+                    num_heads=self.num_heads,
+                    maximum_length=tf.shape(inputs)[1])
+        elif self.self_attention_type == 'average':
+            if cache is None:
+                if sequence_length is None:
+                    sequence_length = tf.fill([tf.shape(inputs)[0]],
+                                              tf.shape(inputs)[1])
+                decoder_mask = transformer.cumulative_average_mask(
+                    sequence_length,
+                    maximum_length=tf.shape(inputs)[1],
+                    dtype=inputs.dtype)
+
+        if memory is not None and not tf.contrib.framework.nest.is_sequence(
+                memory):
+            memory = (memory, )
+        if memory_sequence_length is not None:
+            if not tf.contrib.framework.nest.is_sequence(
+                    memory_sequence_length):
+                memory_sequence_length = (memory_sequence_length, )
+            memory_mask = [
+                transformer.build_sequence_mask(
+                    length,
+                    num_heads=self.num_heads,
+                    maximum_length=tf.shape(m)[1])
+                for m, length in zip(memory, memory_sequence_length)
+            ]
+
+        for layer in range(self.num_layers):
+            layer_name = 'layer_{}'.format(layer)
+            layer_cache = cache[layer_name] if cache is not None else None
+            with tf.variable_scope(layer_name):
+                if self.self_attention_type == 'scaled_dot':
+                    with tf.variable_scope('masked_multi_head'):
+                        encoded = transformer.multi_head_attention(
+                            self.num_heads,
+                            transformer.norm(inputs),
+                            None,
+                            mode,
+                            num_units=self.num_units,
+                            mask=decoder_mask,
+                            cache=layer_cache,
+                            dropout=self.attention_dropout)
+                        last_context = transformer.drop_and_add(
+                            inputs, encoded, mode, dropout=self.dropout)
+                elif self.self_attention_type == 'average':
+                    with tf.variable_scope('average_attention'):
+                        # Cumulative average.
+                        x = transformer.norm(inputs)
+                        y = transformer.cumulative_average(
+                            x,
+                            decoder_mask if cache is None else step,
+                            cache=layer_cache)
+                        # FFN.
+                        y = transformer.feed_forward(
+                            y,
+                            self.ffn_inner_dim,
+                            mode,
+                            dropout=self.relu_dropout)
+                        # Gating layer.
+                        z = tf.layers.dense(
+                            tf.concat([x, y], -1), self.num_units * 2)
+                        i, f = tf.split(z, 2, axis=-1)
+                        y = tf.sigmoid(i) * x + tf.sigmoid(f) * y
+                        last_context = transformer.drop_and_add(
+                            inputs, y, mode, dropout=self.dropout)
+
+                if memory is not None:
+                    for i, (mem, mask) in enumerate(zip(memory, memory_mask)):
+                        memory_cache = layer_cache['memory'][i] if layer_cache is not None else None  # yapf:disable
+                        with tf.variable_scope('multi_head' if i
+                                               == 0 else 'multi_head_%d' % i):  # yapf:disable
+                            context, last_attention = transformer.multi_head_attention(
+                                self.num_heads,
+                                transformer.norm(last_context),
+                                mem,
+                                mode,
+                                mask=mask,
+                                cache=memory_cache,
+                                dropout=self.attention_dropout,
+                                return_attention=True)
+                            last_context = transformer.drop_and_add(
+                                last_context,
+                                context,
+                                mode,
+                                dropout=self.dropout)
+                            if i > 0:  # Do not return attention in case of multi source.
+                                last_attention = None
+
+                with tf.variable_scope('ffn'):
+                    transformed = transformer.feed_forward_ori(
+                        transformer.norm(last_context),
+                        self.ffn_inner_dim,
+                        mode,
+                        dropout=self.relu_dropout)
+                    transformed = transformer.drop_and_add(
+                        last_context, transformed, mode, dropout=self.dropout)
+
+                inputs = transformed
+
+        if last_attention is not None:
+            # The first head of the last layer is returned.
+            first_head_attention = last_attention[:, 0]
+        else:
+            first_head_attention = None
+
+        outputs = transformer.norm(inputs)
+        return outputs, first_head_attention
+
+    def decode_from_inputs(self,
+                           inputs,
+                           sequence_length,
+                           initial_state=None,
+                           mode=True,
+                           memory=None,
+                           memory_sequence_length=None):
+        outputs, attention = self._self_attention_stack(
+            inputs,
+            sequence_length=sequence_length,
+            mode=mode,
+            memory=memory,
+            memory_sequence_length=memory_sequence_length)
+        return outputs, None, attention
+
+    def step_fn(self,
+                mode,
+                batch_size,
+                initial_state=None,
+                memory=None,
+                memory_sequence_length=None,
+                dtype=tf.float32):
+        if memory is None:
+            num_sources = 0
+        elif tf.contrib.framework.nest.is_sequence(memory):
+            num_sources = len(memory)
+        else:
+            num_sources = 1
+        cache = self._init_cache(
+            batch_size, dtype=dtype, num_sources=num_sources)
+
+        def _fn(step, inputs, cache, mode):
+            inputs = tf.expand_dims(inputs, 1)
+            outputs, attention = self._self_attention_stack(
+                inputs,
+                mode=mode,
+                cache=cache,
+                memory=memory,
+                memory_sequence_length=memory_sequence_length,
+                step=step)
+            outputs = tf.squeeze(outputs, axis=1)
+            if attention is not None:
+                attention = tf.squeeze(attention, axis=1)
+            return outputs, cache, attention
+
+        return _fn, cache

modelscope/models/audio/tts/am/models/self_attention_encoder.py

@@ -0,0 +1,182 @@
+"""Define the self-attention encoder."""
+
+import tensorflow as tf
+
+from . import transformer
+from .position import SinusoidalPositionEncoder
+
+
+class SelfAttentionEncoder():
+    """Encoder using self-attention as described in
+    https://arxiv.org/abs/1706.03762.
+    """
+
+    def __init__(self,
+                 num_layers,
+                 num_units=512,
+                 num_heads=8,
+                 ffn_inner_dim=2048,
+                 dropout=0.1,
+                 attention_dropout=0.1,
+                 relu_dropout=0.1,
+                 position_encoder=SinusoidalPositionEncoder()):
+        """Initializes the parameters of the encoder.
+
+        Args:
+          num_layers: The number of layers.
+          num_units: The number of hidden units.
+          num_heads: The number of heads in the multi-head attention.
+          ffn_inner_dim: The number of units of the inner linear transformation
+            in the feed forward layer.
+          dropout: The probability to drop units from the outputs.
+          attention_dropout: The probability to drop units from the attention.
+          relu_dropout: The probability to drop units from the ReLU activation in
+            the feed forward layer.
+          position_encoder: The :class:`opennmt.layers.position.PositionEncoder` to
+            apply on inputs or ``None``.
+        """
+        super(SelfAttentionEncoder, self).__init__()
+        self.num_layers = num_layers
+        self.num_units = num_units
+        self.num_heads = num_heads
+        self.ffn_inner_dim = ffn_inner_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.relu_dropout = relu_dropout
+        self.position_encoder = position_encoder
+
+    def encode(self, inputs, sequence_length=None, mode=True):
+        inputs *= self.num_units**0.5
+        if self.position_encoder is not None:
+            inputs = self.position_encoder(inputs)
+
+        inputs = tf.layers.dropout(inputs, rate=self.dropout, training=mode)
+        mask = transformer.build_sequence_mask(
+            sequence_length,
+            num_heads=self.num_heads,
+            maximum_length=tf.shape(inputs)[1])
+
+        mask_FF = tf.squeeze(
+            transformer.build_sequence_mask(
+                sequence_length, maximum_length=tf.shape(inputs)[1]),
+            axis=1)
+
+        state = ()
+
+        attns = []
+        for layer in range(self.num_layers):
+            with tf.variable_scope('layer_{}'.format(layer)):
+                with tf.variable_scope('multi_head'):
+                    context, attn = transformer.multi_head_attention(
+                        self.num_heads,
+                        transformer.norm(inputs),
+                        None,
+                        mode,
+                        num_units=self.num_units,
+                        mask=mask,
+                        dropout=self.attention_dropout,
+                        return_attention=True)
+                    attns.append(attn)
+                    context = transformer.drop_and_add(
+                        inputs, context, mode, dropout=self.dropout)
+
+                with tf.variable_scope('ffn'):
+                    transformed = transformer.feed_forward(
+                        transformer.norm(context),
+                        self.ffn_inner_dim,
+                        mode,
+                        dropout=self.relu_dropout,
+                        mask=mask_FF)
+                    transformed = transformer.drop_and_add(
+                        context, transformed, mode, dropout=self.dropout)
+
+                inputs = transformed
+                state += (tf.reduce_mean(inputs, axis=1), )
+
+        outputs = transformer.norm(inputs)
+        return (outputs, state, sequence_length, attns)
+
+
+class SelfAttentionEncoderOri():
+    """Encoder using self-attention as described in
+    https://arxiv.org/abs/1706.03762.
+    """
+
+    def __init__(self,
+                 num_layers,
+                 num_units=512,
+                 num_heads=8,
+                 ffn_inner_dim=2048,
+                 dropout=0.1,
+                 attention_dropout=0.1,
+                 relu_dropout=0.1,
+                 position_encoder=SinusoidalPositionEncoder()):
+        """Initializes the parameters of the encoder.
+
+        Args:
+          num_layers: The number of layers.
+          num_units: The number of hidden units.
+          num_heads: The number of heads in the multi-head attention.
+          ffn_inner_dim: The number of units of the inner linear transformation
+            in the feed forward layer.
+          dropout: The probability to drop units from the outputs.
+          attention_dropout: The probability to drop units from the attention.
+          relu_dropout: The probability to drop units from the ReLU activation in
+            the feed forward layer.
+          position_encoder: The :class:`opennmt.layers.position.PositionEncoder` to
+            apply on inputs or ``None``.
+        """
+        super(SelfAttentionEncoderOri, self).__init__()
+        self.num_layers = num_layers
+        self.num_units = num_units
+        self.num_heads = num_heads
+        self.ffn_inner_dim = ffn_inner_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.relu_dropout = relu_dropout
+        self.position_encoder = position_encoder
+
+    def encode(self, inputs, sequence_length=None, mode=True):
+        inputs *= self.num_units**0.5
+        if self.position_encoder is not None:
+            inputs = self.position_encoder(inputs)
+
+        inputs = tf.layers.dropout(inputs, rate=self.dropout, training=mode)
+        mask = transformer.build_sequence_mask(
+            sequence_length,
+            num_heads=self.num_heads,
+            maximum_length=tf.shape(inputs)[1])  # [N, 1, 1, T_out]
+
+        state = ()
+
+        attns = []
+        for layer in range(self.num_layers):
+            with tf.variable_scope('layer_{}'.format(layer)):
+                with tf.variable_scope('multi_head'):
+                    context, attn = transformer.multi_head_attention(
+                        self.num_heads,
+                        transformer.norm(inputs),
+                        None,
+                        mode,
+                        num_units=self.num_units,
+                        mask=mask,
+                        dropout=self.attention_dropout,
+                        return_attention=True)
+                    attns.append(attn)
+                    context = transformer.drop_and_add(
+                        inputs, context, mode, dropout=self.dropout)
+
+                with tf.variable_scope('ffn'):
+                    transformed = transformer.feed_forward_ori(
+                        transformer.norm(context),
+                        self.ffn_inner_dim,
+                        mode,
+                        dropout=self.relu_dropout)
+                    transformed = transformer.drop_and_add(
+                        context, transformed, mode, dropout=self.dropout)
+
+                inputs = transformed
+                state += (tf.reduce_mean(inputs, axis=1), )
+
+        outputs = transformer.norm(inputs)
+        return (outputs, state, sequence_length, attns)

modelscope/models/audio/tts/am/sambert_hifi_16k.py

@@ -0,0 +1,255 @@
+import io
+import os
+from typing import Any, Dict, Optional, Union
+
+import numpy as np
+import tensorflow as tf
+from sklearn.preprocessing import MultiLabelBinarizer
+
+from modelscope.models.base import Model
+from modelscope.models.builder import MODELS
+from modelscope.utils.constant import ModelFile, Tasks
+from .models import create_model
+from .text.symbols import load_symbols
+from .text.symbols_dict import SymbolsDict
+
+__all__ = ['SambertNetHifi16k']
+
+
+def multi_label_symbol_to_sequence(my_classes, my_symbol):
+    one_hot = MultiLabelBinarizer(my_classes)
+    tokens = my_symbol.strip().split(' ')
+    sequences = []
+    for token in tokens:
+        sequences.append(tuple(token.split('&')))
+    # sequences.append(tuple(['~'])) # sequence length minus 1 to ignore EOS ~
+    return one_hot.fit_transform(sequences)
+
+
+@MODELS.register_module(Tasks.text_to_speech, module_name=r'sambert_hifi_16k')
+class SambertNetHifi16k(Model):
+
+    def __init__(self,
+                 model_dir,
+                 pitch_control_str='',
+                 duration_control_str='',
+                 energy_control_str='',
+                 *args,
+                 **kwargs):
+        tf.reset_default_graph()
+        local_ckpt_path = os.path.join(ModelFile.TF_CHECKPOINT_FOLDER, 'ckpt')
+        self._ckpt_path = os.path.join(model_dir, local_ckpt_path)
+        self._dict_path = os.path.join(model_dir, 'dicts')
+        self._hparams = tf.contrib.training.HParams(**kwargs)
+        values = self._hparams.values()
+        hp = [' {}:{}'.format(name, values[name]) for name in sorted(values)]
+        print('Hyperparameters:\n' + '\n'.join(hp))
+        super().__init__(self._ckpt_path, *args, **kwargs)
+        model_name = 'robutrans'
+        self._lfeat_type_list = self._hparams.lfeat_type_list.strip().split(
+            ',')
+        sy, tone, syllable_flag, word_segment, emo_category, speaker = load_symbols(
+            self._dict_path)
+        self._sy = sy
+        self._tone = tone
+        self._syllable_flag = syllable_flag
+        self._word_segment = word_segment
+        self._emo_category = emo_category
+        self._speaker = speaker
+        self._inputs_dim = dict()
+        for lfeat_type in self._lfeat_type_list:
+            if lfeat_type == 'sy':
+                self._inputs_dim[lfeat_type] = len(sy)
+            elif lfeat_type == 'tone':
+                self._inputs_dim[lfeat_type] = len(tone)
+            elif lfeat_type == 'syllable_flag':
+                self._inputs_dim[lfeat_type] = len(syllable_flag)
+            elif lfeat_type == 'word_segment':
+                self._inputs_dim[lfeat_type] = len(word_segment)
+            elif lfeat_type == 'emo_category':
+                self._inputs_dim[lfeat_type] = len(emo_category)
+            elif lfeat_type == 'speaker':
+                self._inputs_dim[lfeat_type] = len(speaker)
+
+        self._symbols_dict = SymbolsDict(sy, tone, syllable_flag, word_segment,
+                                         emo_category, speaker,
+                                         self._inputs_dim,
+                                         self._lfeat_type_list)
+        dim_inputs = sum(self._inputs_dim.values(
+        )) - self._inputs_dim['speaker'] - self._inputs_dim['emo_category']
+        inputs = tf.placeholder(tf.float32, [1, None, dim_inputs], 'inputs')
+        inputs_emotion = tf.placeholder(
+            tf.float32, [1, None, self._inputs_dim['emo_category']],
+            'inputs_emotion')
+        inputs_speaker = tf.placeholder(tf.float32,
+                                        [1, None, self._inputs_dim['speaker']],
+                                        'inputs_speaker')
+
+        input_lengths = tf.placeholder(tf.int32, [1], 'input_lengths')
+        pitch_contours_scale = tf.placeholder(tf.float32, [1, None],
+                                              'pitch_contours_scale')
+        energy_contours_scale = tf.placeholder(tf.float32, [1, None],
+                                               'energy_contours_scale')
+        duration_scale = tf.placeholder(tf.float32, [1, None],
+                                        'duration_scale')
+
+        with tf.variable_scope('model') as _:
+            self._model = create_model(model_name, self._hparams)
+            self._model.initialize(
+                inputs,
+                inputs_emotion,
+                inputs_speaker,
+                input_lengths,
+                duration_scales=duration_scale,
+                pitch_scales=pitch_contours_scale,
+                energy_scales=energy_contours_scale)
+            self._mel_spec = self._model.mel_outputs[0]
+            self._duration_outputs = self._model.duration_outputs[0]
+            self._duration_outputs_ = self._model.duration_outputs_[0]
+            self._pitch_contour_outputs = self._model.pitch_contour_outputs[0]
+            self._energy_contour_outputs = self._model.energy_contour_outputs[
+                0]
+            self._embedded_inputs_emotion = self._model.embedded_inputs_emotion[
+                0]
+            self._embedding_fsmn_outputs = self._model.embedding_fsmn_outputs[
+                0]
+            self._encoder_outputs = self._model.encoder_outputs[0]
+            self._pitch_embeddings = self._model.pitch_embeddings[0]
+            self._energy_embeddings = self._model.energy_embeddings[0]
+            self._LR_outputs = self._model.LR_outputs[0]
+            self._postnet_fsmn_outputs = self._model.postnet_fsmn_outputs[0]
+            self._attention_h = self._model.attention_h
+            self._attention_x = self._model.attention_x
+
+            print('Loading checkpoint: %s' % self._ckpt_path)
+            config = tf.ConfigProto()
+            config.gpu_options.allow_growth = True
+            self._session = tf.Session(config=config)
+            self._session.run(tf.global_variables_initializer())
+
+            saver = tf.train.Saver()
+            saver.restore(self._session, self._ckpt_path)
+
+            duration_cfg_lst = []
+            if len(duration_control_str) != 0:
+                for item in duration_control_str.strip().split('|'):
+                    percent, scale = item.lstrip('(').rstrip(')').split(',')
+                    duration_cfg_lst.append((float(percent), float(scale)))
+
+            self._duration_cfg_lst = duration_cfg_lst
+
+            pitch_contours_cfg_lst = []
+            if len(pitch_control_str) != 0:
+                for item in pitch_control_str.strip().split('|'):
+                    percent, scale = item.lstrip('(').rstrip(')').split(',')
+                    pitch_contours_cfg_lst.append(
+                        (float(percent), float(scale)))
+
+            self._pitch_contours_cfg_lst = pitch_contours_cfg_lst
+
+            energy_contours_cfg_lst = []
+            if len(energy_control_str) != 0:
+                for item in energy_control_str.strip().split('|'):
+                    percent, scale = item.lstrip('(').rstrip(')').split(',')
+                    energy_contours_cfg_lst.append(
+                        (float(percent), float(scale)))
+
+            self._energy_contours_cfg_lst = energy_contours_cfg_lst
+
+    def forward(self, text):
+        cleaner_names = [x.strip() for x in self._hparams.cleaners.split(',')]
+
+        lfeat_symbol = text.strip().split(' ')
+        lfeat_symbol_separate = [''] * int(len(self._lfeat_type_list))
+        for this_lfeat_symbol in lfeat_symbol:
+            this_lfeat_symbol = this_lfeat_symbol.strip('{').strip('}').split(
+                '$')
+            if len(this_lfeat_symbol) != len(self._lfeat_type_list):
+                raise Exception(
+                    'Length of this_lfeat_symbol in training data'
+                    + ' is not equal to the length of lfeat_type_list, '
+                    + str(len(this_lfeat_symbol)) + ' VS. '
+                    + str(len(self._lfeat_type_list)))
+            index = 0
+            while index < len(lfeat_symbol_separate):
+                lfeat_symbol_separate[index] = lfeat_symbol_separate[
+                    index] + this_lfeat_symbol[index] + ' '
+                index = index + 1
+
+        index = 0
+        lfeat_type = self._lfeat_type_list[index]
+        sequence = self._symbols_dict.symbol_to_sequence(
+            lfeat_symbol_separate[index].strip(), lfeat_type, cleaner_names)
+        sequence_array = np.asarray(
+            sequence[:-1],
+            dtype=np.int32)  # sequence length minus 1 to ignore EOS ~
+        inputs = np.eye(
+            self._inputs_dim[lfeat_type], dtype=np.float32)[sequence_array]
+        index = index + 1
+        while index < len(self._lfeat_type_list) - 2:
+            lfeat_type = self._lfeat_type_list[index]
+            sequence = self._symbols_dict.symbol_to_sequence(
+                lfeat_symbol_separate[index].strip(), lfeat_type,
+                cleaner_names)
+            sequence_array = np.asarray(
+                sequence[:-1],
+                dtype=np.int32)  # sequence length minus 1 to ignore EOS ~
+            inputs_temp = np.eye(
+                self._inputs_dim[lfeat_type], dtype=np.float32)[sequence_array]
+            inputs = np.concatenate((inputs, inputs_temp), axis=1)
+            index = index + 1
+        seq = inputs
+
+        lfeat_type = 'emo_category'
+        inputs_emotion = multi_label_symbol_to_sequence(
+            self._emo_category, lfeat_symbol_separate[index].strip())
+        # inputs_emotion = inputs_emotion * 1.5
+        index = index + 1
+
+        lfeat_type = 'speaker'
+        inputs_speaker = multi_label_symbol_to_sequence(
+            self._speaker, lfeat_symbol_separate[index].strip())
+
+        duration_scale = np.ones((len(seq), ), dtype=np.float32)
+        start_idx = 0
+        for (percent, scale) in self._duration_cfg_lst:
+            duration_scale[start_idx:start_idx
+                           + int(percent * len(seq))] = scale
+            start_idx += int(percent * len(seq))
+
+        pitch_contours_scale = np.ones((len(seq), ), dtype=np.float32)
+        start_idx = 0
+        for (percent, scale) in self._pitch_contours_cfg_lst:
+            pitch_contours_scale[start_idx:start_idx
+                                 + int(percent * len(seq))] = scale
+            start_idx += int(percent * len(seq))
+
+        energy_contours_scale = np.ones((len(seq), ), dtype=np.float32)
+        start_idx = 0
+        for (percent, scale) in self._energy_contours_cfg_lst:
+            energy_contours_scale[start_idx:start_idx
+                                  + int(percent * len(seq))] = scale
+            start_idx += int(percent * len(seq))
+
+        feed_dict = {
+            self._model.inputs: [np.asarray(seq, dtype=np.float32)],
+            self._model.inputs_emotion:
+            [np.asarray(inputs_emotion, dtype=np.float32)],
+            self._model.inputs_speaker:
+            [np.asarray(inputs_speaker, dtype=np.float32)],
+            self._model.input_lengths:
+            np.asarray([len(seq)], dtype=np.int32),
+            self._model.duration_scales: [duration_scale],
+            self._model.pitch_scales: [pitch_contours_scale],
+            self._model.energy_scales: [energy_contours_scale]
+        }
+
+        result = self._session.run([
+            self._mel_spec, self._duration_outputs, self._duration_outputs_,
+            self._pitch_contour_outputs, self._embedded_inputs_emotion,
+            self._embedding_fsmn_outputs, self._encoder_outputs,
+            self._pitch_embeddings, self._LR_outputs,
+            self._postnet_fsmn_outputs, self._energy_contour_outputs,
+            self._energy_embeddings, self._attention_x, self._attention_h
+        ], feed_dict=feed_dict)  # yapf:disable
+        return result[0]

modelscope/models/audio/tts/am/text/cleaners.py

@@ -0,0 +1,89 @@
+'''
+Cleaners are transformations that run over the input text at both training and eval time.
+
+Cleaners can be selected by passing a comma-delimited list of cleaner names as the "cleaners"
+hyperparameter. Some cleaners are English-specific. You'll typically want to use:
+  1. "english_cleaners" for English text
+  2. "transliteration_cleaners" for non-English text that can be transliterated to ASCII using
+     the Unidecode library (https://pypi.python.org/pypi/Unidecode)
+  3. "basic_cleaners" if you do not want to transliterate (in this case, you should also update
+     the symbols in symbols.py to match your data).
+'''
+
+import re
+
+from unidecode import unidecode
+
+from .numbers import normalize_numbers
+
+# Regular expression matching whitespace:
+_whitespace_re = re.compile(r'\s+')
+
+# List of (regular expression, replacement) pairs for abbreviations:
+_abbreviations = [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1])
+                  for x in [
+                      ('mrs', 'misess'),
+                      ('mr', 'mister'),
+                      ('dr', 'doctor'),
+                      ('st', 'saint'),
+                      ('co', 'company'),
+                      ('jr', 'junior'),
+                      ('maj', 'major'),
+                      ('gen', 'general'),
+                      ('drs', 'doctors'),
+                      ('rev', 'reverend'),
+                      ('lt', 'lieutenant'),
+                      ('hon', 'honorable'),
+                      ('sgt', 'sergeant'),
+                      ('capt', 'captain'),
+                      ('esq', 'esquire'),
+                      ('ltd', 'limited'),
+                      ('col', 'colonel'),
+                      ('ft', 'fort'), ]]  # yapf:disable
+
+
+def expand_abbreviations(text):
+    for regex, replacement in _abbreviations:
+        text = re.sub(regex, replacement, text)
+    return text
+
+
+def expand_numbers(text):
+    return normalize_numbers(text)
+
+
+def lowercase(text):
+    return text.lower()
+
+
+def collapse_whitespace(text):
+    return re.sub(_whitespace_re, ' ', text)
+
+
+def convert_to_ascii(text):
+    return unidecode(text)
+
+
+def basic_cleaners(text):
+    '''Basic pipeline that lowercases and collapses whitespace without transliteration.'''
+    text = lowercase(text)
+    text = collapse_whitespace(text)
+    return text
+
+
+def transliteration_cleaners(text):
+    '''Pipeline for non-English text that transliterates to ASCII.'''
+    text = convert_to_ascii(text)
+    text = lowercase(text)
+    text = collapse_whitespace(text)
+    return text
+
+
+def english_cleaners(text):
+    '''Pipeline for English text, including number and abbreviation expansion.'''
+    text = convert_to_ascii(text)
+    text = lowercase(text)
+    text = expand_numbers(text)
+    text = expand_abbreviations(text)
+    text = collapse_whitespace(text)
+    return text

modelscope/models/audio/tts/am/text/cmudict.py

@@ -0,0 +1,64 @@
+import re
+
+valid_symbols = [
+    'AA', 'AA0', 'AA1', 'AA2', 'AE', 'AE0', 'AE1', 'AE2', 'AH', 'AH0', 'AH1',
+    'AH2', 'AO', 'AO0', 'AO1', 'AO2', 'AW', 'AW0', 'AW1', 'AW2', 'AY', 'AY0',
+    'AY1', 'AY2', 'B', 'CH', 'D', 'DH', 'EH', 'EH0', 'EH1', 'EH2', 'ER', 'ER0',
+    'ER1', 'ER2', 'EY', 'EY0', 'EY1', 'EY2', 'F', 'G', 'HH', 'IH', 'IH0',
+    'IH1', 'IH2', 'IY', 'IY0', 'IY1', 'IY2', 'JH', 'K', 'L', 'M', 'N', 'NG',
+    'OW', 'OW0', 'OW1', 'OW2', 'OY', 'OY0', 'OY1', 'OY2', 'P', 'R', 'S', 'SH',
+    'T', 'TH', 'UH', 'UH0', 'UH1', 'UH2', 'UW', 'UW0', 'UW1', 'UW2', 'V', 'W',
+    'Y', 'Z', 'ZH'
+]
+
+_valid_symbol_set = set(valid_symbols)
+
+
+class CMUDict:
+    '''Thin wrapper around CMUDict data. http://www.speech.cs.cmu.edu/cgi-bin/cmudict'''
+
+    def __init__(self, file_or_path, keep_ambiguous=True):
+        if isinstance(file_or_path, str):
+            with open(file_or_path, encoding='latin-1') as f:
+                entries = _parse_cmudict(f)
+        else:
+            entries = _parse_cmudict(file_or_path)
+        if not keep_ambiguous:
+            entries = {
+                word: pron
+                for word, pron in entries.items() if len(pron) == 1
+            }
+        self._entries = entries
+
+    def __len__(self):
+        return len(self._entries)
+
+    def lookup(self, word):
+        '''Returns list of ARPAbet pronunciations of the given word.'''
+        return self._entries.get(word.upper())
+
+
+_alt_re = re.compile(r'\([0-9]+\)')
+
+
+def _parse_cmudict(file):
+    cmudict = {}
+    for line in file:
+        if len(line) and (line[0] >= 'A' and line[0] <= 'Z' or line[0] == "'"):
+            parts = line.split('  ')
+            word = re.sub(_alt_re, '', parts[0])
+            pronunciation = _get_pronunciation(parts[1])
+            if pronunciation:
+                if word in cmudict:
+                    cmudict[word].append(pronunciation)
+                else:
+                    cmudict[word] = [pronunciation]
+    return cmudict
+
+
+def _get_pronunciation(s):
+    parts = s.strip().split(' ')
+    for part in parts:
+        if part not in _valid_symbol_set:
+            return None
+    return ' '.join(parts)

modelscope/models/audio/tts/am/text/numbers.py

@@ -0,0 +1,70 @@
+import re
+
+import inflect
+
+_inflect = inflect.engine()
+_comma_number_re = re.compile(r'([0-9][0-9\,]+[0-9])')
+_decimal_number_re = re.compile(r'([0-9]+\.[0-9]+)')
+_pounds_re = re.compile(r'£([0-9\,]*[0-9]+)')
+_dollars_re = re.compile(r'\$([0-9\.\,]*[0-9]+)')
+_ordinal_re = re.compile(r'[0-9]+(st|nd|rd|th)')
+_number_re = re.compile(r'[0-9]+')
+
+
+def _remove_commas(m):
+    return m.group(1).replace(',', '')
+
+
+def _expand_decimal_point(m):
+    return m.group(1).replace('.', ' point ')
+
+
+def _expand_dollars(m):
+    match = m.group(1)
+    parts = match.split('.')
+    if len(parts) > 2:
+        return match + ' dollars'  # Unexpected format
+    dollars = int(parts[0]) if parts[0] else 0
+    cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
+    if dollars and cents:
+        dollar_unit = 'dollar' if dollars == 1 else 'dollars'
+        cent_unit = 'cent' if cents == 1 else 'cents'
+        return '%s %s, %s %s' % (dollars, dollar_unit, cents, cent_unit)
+    elif dollars:
+        dollar_unit = 'dollar' if dollars == 1 else 'dollars'
+        return '%s %s' % (dollars, dollar_unit)
+    elif cents:
+        cent_unit = 'cent' if cents == 1 else 'cents'
+        return '%s %s' % (cents, cent_unit)
+    else:
+        return 'zero dollars'
+
+
+def _expand_ordinal(m):
+    return _inflect.number_to_words(m.group(0))
+
+
+def _expand_number(m):
+    num = int(m.group(0))
+    if num > 1000 and num < 3000:
+        if num == 2000:
+            return 'two thousand'
+        elif num > 2000 and num < 2010:
+            return 'two thousand ' + _inflect.number_to_words(num % 100)
+        elif num % 100 == 0:
+            return _inflect.number_to_words(num // 100) + ' hundred'
+        else:
+            return _inflect.number_to_words(
+                num, andword='', zero='oh', group=2).replace(', ', ' ')
+    else:
+        return _inflect.number_to_words(num, andword='')
+
+
+def normalize_numbers(text):
+    text = re.sub(_comma_number_re, _remove_commas, text)
+    text = re.sub(_pounds_re, r'\1 pounds', text)
+    text = re.sub(_dollars_re, _expand_dollars, text)
+    text = re.sub(_decimal_number_re, _expand_decimal_point, text)
+    text = re.sub(_ordinal_re, _expand_ordinal, text)
+    text = re.sub(_number_re, _expand_number, text)
+    return text

modelscope/models/audio/tts/am/text/symbols.py

@@ -0,0 +1,95 @@
+'''
+Defines the set of symbols used in text input to the model.
+
+The default is a set of ASCII characters that works well for English or text that has been run
+through Unidecode. For other data, you can modify _characters. See TRAINING_DATA.md for details.
+'''
+import codecs
+import os
+
+_pad = '_'
+_eos = '~'
+_mask = '@[MASK]'
+
+
+def load_symbols(dict_path):
+    _characters = ''
+    _ch_symbols = []
+    sy_dict_name = 'sy_dict.txt'
+    sy_dict_path = os.path.join(dict_path, sy_dict_name)
+    f = codecs.open(sy_dict_path, 'r')
+    for line in f:
+        line = line.strip('\r\n')
+        _ch_symbols.append(line)
+
+    _arpabet = ['@' + s for s in _ch_symbols]
+
+    # Export all symbols:
+    sy = list(_characters) + _arpabet + [_pad, _eos, _mask]
+
+    _characters = ''
+
+    _ch_tones = []
+    tone_dict_name = 'tone_dict.txt'
+    tone_dict_path = os.path.join(dict_path, tone_dict_name)
+    f = codecs.open(tone_dict_path, 'r')
+    for line in f:
+        line = line.strip('\r\n')
+        _ch_tones.append(line)
+
+    # Export all tones:
+    tone = list(_characters) + _ch_tones + [_pad, _eos, _mask]
+
+    _characters = ''
+
+    _ch_syllable_flags = []
+    syllable_flag_name = 'syllable_flag_dict.txt'
+    syllable_flag_path = os.path.join(dict_path, syllable_flag_name)
+    f = codecs.open(syllable_flag_path, 'r')
+    for line in f:
+        line = line.strip('\r\n')
+        _ch_syllable_flags.append(line)
+
+    # Export all syllable_flags:
+    syllable_flag = list(_characters) + _ch_syllable_flags + [
+        _pad, _eos, _mask
+    ]
+
+    _characters = ''
+
+    _ch_word_segments = []
+    word_segment_name = 'word_segment_dict.txt'
+    word_segment_path = os.path.join(dict_path, word_segment_name)
+    f = codecs.open(word_segment_path, 'r')
+    for line in f:
+        line = line.strip('\r\n')
+        _ch_word_segments.append(line)
+
+    # Export all syllable_flags:
+    word_segment = list(_characters) + _ch_word_segments + [_pad, _eos, _mask]
+
+    _characters = ''
+
+    _ch_emo_types = []
+    emo_category_name = 'emo_category_dict.txt'
+    emo_category_path = os.path.join(dict_path, emo_category_name)
+    f = codecs.open(emo_category_path, 'r')
+    for line in f:
+        line = line.strip('\r\n')
+        _ch_emo_types.append(line)
+
+    emo_category = list(_characters) + _ch_emo_types + [_pad, _eos, _mask]
+
+    _characters = ''
+
+    _ch_speakers = []
+    speaker_name = 'speaker_dict.txt'
+    speaker_path = os.path.join(dict_path, speaker_name)
+    f = codecs.open(speaker_path, 'r')
+    for line in f:
+        line = line.strip('\r\n')
+        _ch_speakers.append(line)
+
+    # Export all syllable_flags:
+    speaker = list(_characters) + _ch_speakers + [_pad, _eos, _mask]
+    return sy, tone, syllable_flag, word_segment, emo_category, speaker

modelscope/models/audio/tts/am/text/symbols_dict.py

@@ -0,0 +1,200 @@
+import re
+import sys
+
+from .cleaners import (basic_cleaners, english_cleaners,
+                       transliteration_cleaners)
+
+
+class SymbolsDict:
+
+    def __init__(self, sy, tone, syllable_flag, word_segment, emo_category,
+                 speaker, inputs_dim, lfeat_type_list):
+        self._inputs_dim = inputs_dim
+        self._lfeat_type_list = lfeat_type_list
+        self._sy_to_id = {s: i for i, s in enumerate(sy)}
+        self._id_to_sy = {i: s for i, s in enumerate(sy)}
+        self._tone_to_id = {s: i for i, s in enumerate(tone)}
+        self._id_to_tone = {i: s for i, s in enumerate(tone)}
+        self._syllable_flag_to_id = {s: i for i, s in enumerate(syllable_flag)}
+        self._id_to_syllable_flag = {i: s for i, s in enumerate(syllable_flag)}
+        self._word_segment_to_id = {s: i for i, s in enumerate(word_segment)}
+        self._id_to_word_segment = {i: s for i, s in enumerate(word_segment)}
+        self._emo_category_to_id = {s: i for i, s in enumerate(emo_category)}
+        self._id_to_emo_category = {i: s for i, s in enumerate(emo_category)}
+        self._speaker_to_id = {s: i for i, s in enumerate(speaker)}
+        self._id_to_speaker = {i: s for i, s in enumerate(speaker)}
+        print('_sy_to_id: ')
+        print(self._sy_to_id)
+        print('_tone_to_id: ')
+        print(self._tone_to_id)
+        print('_syllable_flag_to_id: ')
+        print(self._syllable_flag_to_id)
+        print('_word_segment_to_id: ')
+        print(self._word_segment_to_id)
+        print('_emo_category_to_id: ')
+        print(self._emo_category_to_id)
+        print('_speaker_to_id: ')
+        print(self._speaker_to_id)
+        self._curly_re = re.compile(r'(.*?)\{(.+?)\}(.*)')
+        self._cleaners = {
+            basic_cleaners.__name__: basic_cleaners,
+            transliteration_cleaners.__name__: transliteration_cleaners,
+            english_cleaners.__name__: english_cleaners
+        }
+
+    def _clean_text(self, text, cleaner_names):
+        for name in cleaner_names:
+            cleaner = self._cleaners.get(name)
+            if not cleaner:
+                raise Exception('Unknown cleaner: %s' % name)
+            text = cleaner(text)
+        return text
+
+    def _sy_to_sequence(self, sy):
+        return [self._sy_to_id[s] for s in sy if self._should_keep_sy(s)]
+
+    def _arpabet_to_sequence(self, text):
+        return self._sy_to_sequence(['@' + s for s in text.split()])
+
+    def _should_keep_sy(self, s):
+        return s in self._sy_to_id and s != '_' and s != '~'
+
+    def symbol_to_sequence(self, this_lfeat_symbol, lfeat_type, cleaner_names):
+        sequence = []
+        if lfeat_type == 'sy':
+            this_lfeat_symbol = this_lfeat_symbol.strip().split(' ')
+            this_lfeat_symbol_format = ''
+            index = 0
+            while index < len(this_lfeat_symbol):
+                this_lfeat_symbol_format = this_lfeat_symbol_format + '{' + this_lfeat_symbol[
+                    index] + '}' + ' '
+                index = index + 1
+            sequence = self.text_to_sequence(this_lfeat_symbol_format,
+                                             cleaner_names)
+        elif lfeat_type == 'tone':
+            sequence = self.tone_to_sequence(this_lfeat_symbol)
+        elif lfeat_type == 'syllable_flag':
+            sequence = self.syllable_flag_to_sequence(this_lfeat_symbol)
+        elif lfeat_type == 'word_segment':
+            sequence = self.word_segment_to_sequence(this_lfeat_symbol)
+        elif lfeat_type == 'emo_category':
+            sequence = self.emo_category_to_sequence(this_lfeat_symbol)
+        elif lfeat_type == 'speaker':
+            sequence = self.speaker_to_sequence(this_lfeat_symbol)
+        else:
+            raise Exception('Unknown lfeat type: %s' % lfeat_type)
+
+        return sequence
+
+    def text_to_sequence(self, text, cleaner_names):
+        '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+
+          The text can optionally have ARPAbet sequences enclosed in curly braces embedded
+          in it. For example, "Turn left on {HH AW1 S S T AH0 N} Street."
+
+          Args:
+            text: string to convert to a sequence
+            cleaner_names: names of the cleaner functions to run the text through
+
+          Returns:
+            List of integers corresponding to the symbols in the text
+        '''
+        sequence = []
+
+        # Check for curly braces and treat their contents as ARPAbet:
+        while len(text):
+            m = self._curly_re.match(text)
+            if not m:
+                sequence += self._sy_to_sequence(
+                    self._clean_text(text, cleaner_names))
+                break
+            sequence += self._sy_to_sequence(
+                self._clean_text(m.group(1), cleaner_names))
+            sequence += self._arpabet_to_sequence(m.group(2))
+            text = m.group(3)
+
+        # Append EOS token
+        sequence.append(self._sy_to_id['~'])
+        return sequence
+
+    def tone_to_sequence(self, tone):
+        tones = tone.strip().split(' ')
+        sequence = []
+        for this_tone in tones:
+            sequence.append(self._tone_to_id[this_tone])
+        sequence.append(self._tone_to_id['~'])
+        return sequence
+
+    def syllable_flag_to_sequence(self, syllable_flag):
+        syllable_flags = syllable_flag.strip().split(' ')
+        sequence = []
+        for this_syllable_flag in syllable_flags:
+            sequence.append(self._syllable_flag_to_id[this_syllable_flag])
+        sequence.append(self._syllable_flag_to_id['~'])
+        return sequence
+
+    def word_segment_to_sequence(self, word_segment):
+        word_segments = word_segment.strip().split(' ')
+        sequence = []
+        for this_word_segment in word_segments:
+            sequence.append(self._word_segment_to_id[this_word_segment])
+        sequence.append(self._word_segment_to_id['~'])
+        return sequence
+
+    def emo_category_to_sequence(self, emo_type):
+        emo_categories = emo_type.strip().split(' ')
+        sequence = []
+        for this_category in emo_categories:
+            sequence.append(self._emo_category_to_id[this_category])
+        sequence.append(self._emo_category_to_id['~'])
+        return sequence
+
+    def speaker_to_sequence(self, speaker):
+        speakers = speaker.strip().split(' ')
+        sequence = []
+        for this_speaker in speakers:
+            sequence.append(self._speaker_to_id[this_speaker])
+        sequence.append(self._speaker_to_id['~'])
+        return sequence
+
+    def sequence_to_symbol(self, sequence):
+        result = ''
+        pre_lfeat_dim = 0
+        for lfeat_type in self._lfeat_type_list:
+            current_one_hot_sequence = sequence[:, pre_lfeat_dim:pre_lfeat_dim
+                                                + self._inputs_dim[lfeat_type]]
+            current_sequence = current_one_hot_sequence.argmax(1)
+            length = current_sequence.shape[0]
+
+            index = 0
+            while index < length:
+                this_sequence = current_sequence[index]
+                s = ''
+                if lfeat_type == 'sy':
+                    s = self._id_to_sy[this_sequence]
+                    if len(s) > 1 and s[0] == '@':
+                        s = s[1:]
+                elif lfeat_type == 'tone':
+                    s = self._id_to_tone[this_sequence]
+                elif lfeat_type == 'syllable_flag':
+                    s = self._id_to_syllable_flag[this_sequence]
+                elif lfeat_type == 'word_segment':
+                    s = self._id_to_word_segment[this_sequence]
+                elif lfeat_type == 'emo_category':
+                    s = self._id_to_emo_category[this_sequence]
+                elif lfeat_type == 'speaker':
+                    s = self._id_to_speaker[this_sequence]
+                else:
+                    raise Exception('Unknown lfeat type: %s' % lfeat_type)
+
+                if index == 0:
+                    result = result + lfeat_type + ': '
+
+                result = result + '{' + s + '}'
+
+                if index == length - 1:
+                    result = result + '; '
+
+                index = index + 1
+            pre_lfeat_dim = pre_lfeat_dim + self._inputs_dim[lfeat_type]
+        return result

modelscope/models/audio/tts/frontend/__init__.py

@@ -0,0 +1 @@
+from .generic_text_to_speech_frontend import *  # noqa F403

modelscope/models/audio/tts/frontend/generic_text_to_speech_frontend.py

@@ -0,0 +1,39 @@
+import os
+import zipfile
+from typing import Any, Dict, List
+
+from modelscope.models.base import Model
+from modelscope.models.builder import MODELS
+from modelscope.utils.audio.tts_exceptions import (
+    TtsFrontendInitializeFailedException,
+    TtsFrontendLanguageTypeInvalidException)
+from modelscope.utils.constant import Tasks
+
+__all__ = ['GenericTtsFrontend']
+
+
+@MODELS.register_module(
+    Tasks.text_to_speech, module_name=r'generic_tts_frontend')
+class GenericTtsFrontend(Model):
+
+    def __init__(self, model_dir='.', lang_type='pinyin', *args, **kwargs):
+        super().__init__(model_dir, *args, **kwargs)
+        import ttsfrd
+
+        frontend = ttsfrd.TtsFrontendEngine()
+        zip_file = os.path.join(model_dir, 'resource.zip')
+        self._res_path = os.path.join(model_dir, 'resource')
+        with zipfile.ZipFile(zip_file, 'r') as zip_ref:
+            zip_ref.extractall(model_dir)
+        if not frontend.initialize(self._res_path):
+            raise TtsFrontendInitializeFailedException(
+                'resource invalid: {}'.format(self._res_path))
+        if not frontend.set_lang_type(lang_type):
+            raise TtsFrontendLanguageTypeInvalidException(
+                'language type invalid: {}, valid is pinyin and chenmix'.
+                format(lang_type))
+        self._frontend = frontend
+
+    def forward(self, data: str) -> Dict[str, List]:
+        result = self._frontend.gen_tacotron_symbols(data)
+        return {'texts': [s for s in result.splitlines() if s != '']}

modelscope/models/audio/tts/vocoder/__init__.py

@@ -0,0 +1 @@
+from .hifigan16k import *  # noqa F403

modelscope/models/audio/tts/vocoder/hifigan16k.py

@@ -0,0 +1,73 @@
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+import argparse
+import glob
+import os
+import time
+
+import json
+import numpy as np
+import torch
+from scipy.io.wavfile import write
+
+from modelscope.models.base import Model
+from modelscope.models.builder import MODELS
+from modelscope.utils.audio.tts_exceptions import \
+    TtsVocoderMelspecShapeMismatchException
+from modelscope.utils.constant import ModelFile, Tasks
+from .models import Generator
+
+__all__ = ['Hifigan16k', 'AttrDict']
+MAX_WAV_VALUE = 32768.0
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print('Complete.')
+    return checkpoint_dict
+
+
+class AttrDict(dict):
+
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+@MODELS.register_module(Tasks.text_to_speech, module_name=r'hifigan16k')
+class Hifigan16k(Model):
+
+    def __init__(self, model_dir, *args, **kwargs):
+        self._ckpt_path = os.path.join(model_dir,
+                                       ModelFile.TORCH_MODEL_BIN_FILE)
+        self._config = AttrDict(**kwargs)
+
+        super().__init__(self._ckpt_path, *args, **kwargs)
+        if torch.cuda.is_available():
+            torch.manual_seed(self._config.seed)
+            self._device = torch.device('cuda')
+        else:
+            self._device = torch.device('cpu')
+        self._generator = Generator(self._config).to(self._device)
+        state_dict_g = load_checkpoint(self._ckpt_path, self._device)
+        self._generator.load_state_dict(state_dict_g['generator'])
+        self._generator.eval()
+        self._generator.remove_weight_norm()
+
+    def forward(self, melspec):
+        dim0 = list(melspec.shape)[-1]
+        if dim0 != 80:
+            raise TtsVocoderMelspecShapeMismatchException(
+                'input melspec mismatch 0 dim require 80 but {}'.format(dim0))
+        with torch.no_grad():
+            x = melspec.T
+            x = torch.FloatTensor(x).to(self._device)
+            if len(x.shape) == 2:
+                x = x.unsqueeze(0)
+            y_g_hat = self._generator(x)
+            audio = y_g_hat.squeeze()
+            audio = audio * MAX_WAV_VALUE
+            audio = audio.cpu().numpy().astype('int16')
+            return audio

modelscope/models/audio/tts/vocoder/models/__init__.py

@@ -0,0 +1 @@
+from .models import Generator

modelscope/models/audio/tts/vocoder/models/models.py

@@ -0,0 +1,516 @@
+from distutils.version import LooseVersion
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
+from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
+
+from .utils import get_padding, init_weights
+
+is_pytorch_17plus = LooseVersion(torch.__version__) >= LooseVersion('1.7')
+
+
+def stft(x, fft_size, hop_size, win_length, window):
+    """Perform STFT and convert to magnitude spectrogram.
+
+    Args:
+        x (Tensor): Input signal tensor (B, T).
+        fft_size (int): FFT size.
+        hop_size (int): Hop size.
+        win_length (int): Window length.
+        window (str): Window function type.
+
+    Returns:
+        Tensor: Magnitude spectrogram (B, #frames, fft_size // 2 + 1).
+
+    """
+    if is_pytorch_17plus:
+        x_stft = torch.stft(
+            x, fft_size, hop_size, win_length, window, return_complex=False)
+    else:
+        x_stft = torch.stft(x, fft_size, hop_size, win_length, window)
+    real = x_stft[..., 0]
+    imag = x_stft[..., 1]
+
+    # NOTE(kan-bayashi): clamp is needed to avoid nan or inf
+    return torch.sqrt(torch.clamp(real**2 + imag**2, min=1e-7)).transpose(2, 1)
+
+
+LRELU_SLOPE = 0.1
+
+
+def get_padding_casual(kernel_size, dilation=1):
+    return int(kernel_size * dilation - dilation)
+
+
+class Conv1dCasual(torch.nn.Module):
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1,
+                 bias=True,
+                 padding_mode='zeros'):
+        super(Conv1dCasual, self).__init__()
+        self.pad = padding
+        self.conv1d = weight_norm(
+            Conv1d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                stride,
+                padding=0,
+                dilation=dilation,
+                groups=groups,
+                bias=bias,
+                padding_mode=padding_mode))
+        self.conv1d.apply(init_weights)
+
+    def forward(self, x):  # bdt
+        # described starting from the last dimension and moving forward.
+        x = F.pad(x, (self.pad, 0, 0, 0, 0, 0), 'constant')
+        x = self.conv1d(x)
+        return x
+
+    def remove_weight_norm(self):
+        remove_weight_norm(self.conv1d)
+
+
+class ConvTranspose1dCausal(torch.nn.Module):
+    """CausalConvTranspose1d module with customized initialization."""
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 padding=0):
+        """Initialize CausalConvTranspose1d module."""
+        super(ConvTranspose1dCausal, self).__init__()
+        self.deconv = weight_norm(
+            ConvTranspose1d(in_channels, out_channels, kernel_size, stride))
+        self.stride = stride
+        self.deconv.apply(init_weights)
+        self.pad = kernel_size - stride
+
+    def forward(self, x):
+        """Calculate forward propagation.
+        Args:
+            x (Tensor): Input tensor (B, in_channels, T_in).
+        Returns:
+            Tensor: Output tensor (B, out_channels, T_out).
+        """
+        # x = F.pad(x, (self.pad, 0, 0, 0, 0, 0), "constant")
+        return self.deconv(x)[:, :, :-self.pad]
+
+    def remove_weight_norm(self):
+        remove_weight_norm(self.deconv)
+
+
+class ResBlock1(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList([
+            Conv1dCasual(
+                channels,
+                channels,
+                kernel_size,
+                1,
+                dilation=dilation[i],
+                padding=get_padding_casual(kernel_size, dilation[i]))
+            for i in range(len(dilation))
+        ])
+
+        self.convs2 = nn.ModuleList([
+            Conv1dCasual(
+                channels,
+                channels,
+                kernel_size,
+                1,
+                dilation=1,
+                padding=get_padding_casual(kernel_size, 1))
+            for i in range(len(dilation))
+        ])
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for layer in self.convs1:
+            layer.remove_weight_norm()
+        for layer in self.convs2:
+            layer.remove_weight_norm()
+
+
+class Generator(torch.nn.Module):
+
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        print('num_kernels={}, num_upsamples={}'.format(
+            self.num_kernels, self.num_upsamples))
+        self.conv_pre = Conv1dCasual(
+            80, h.upsample_initial_channel, 7, 1, padding=7 - 1)
+        resblock = ResBlock1 if h.resblock == '1' else ResBlock2
+
+        self.ups = nn.ModuleList()
+        self.repeat_ups = nn.ModuleList()
+        for i, (u, k) in enumerate(
+                zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            upsample = nn.Sequential(
+                nn.Upsample(mode='nearest', scale_factor=u),
+                nn.LeakyReLU(LRELU_SLOPE),
+                Conv1dCasual(
+                    h.upsample_initial_channel // (2**i),
+                    h.upsample_initial_channel // (2**(i + 1)),
+                    kernel_size=7,
+                    stride=1,
+                    padding=7 - 1))
+            self.repeat_ups.append(upsample)
+            self.ups.append(
+                ConvTranspose1dCausal(
+                    h.upsample_initial_channel // (2**i),
+                    h.upsample_initial_channel // (2**(i + 1)),
+                    k,
+                    u,
+                    padding=(k - u) // 2))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2**(i + 1))
+            for j, (k, d) in enumerate(
+                    zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = Conv1dCasual(ch, 1, 7, 1, padding=7 - 1)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = torch.sin(x) + x
+            # transconv
+            x1 = F.leaky_relu(x, LRELU_SLOPE)
+            x1 = self.ups[i](x1)
+            # repeat
+            x2 = self.repeat_ups[i](x)
+            x = x1 + x2
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for layer in self.ups:
+            layer.remove_weight_norm()
+        for layer in self.repeat_ups:
+            layer[-1].remove_weight_norm()
+        for layer in self.resblocks:
+            layer.remove_weight_norm()
+        self.conv_pre.remove_weight_norm()
+        self.conv_post.remove_weight_norm()
+
+
+class DiscriminatorP(torch.nn.Module):
+
+    def __init__(self,
+                 period,
+                 kernel_size=5,
+                 stride=3,
+                 use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(
+                Conv2d(
+                    1,
+                    32, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(5, 1), 0))),
+            norm_f(
+                Conv2d(
+                    32,
+                    128, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(5, 1), 0))),
+            norm_f(
+                Conv2d(
+                    128,
+                    512, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(5, 1), 0))),
+            norm_f(
+                Conv2d(
+                    512,
+                    1024, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(5, 1), 0))),
+            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
+        ])
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), 'reflect')
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+
+    def __init__(self):
+        super(MultiPeriodDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList([
+            DiscriminatorP(2),
+            DiscriminatorP(3),
+            DiscriminatorP(5),
+            DiscriminatorP(7),
+            DiscriminatorP(11),
+        ])
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(Conv1d(1, 128, 15, 1, padding=7)),
+            norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)),
+            norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)),
+            norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)),
+            norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
+            norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
+            norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+        ])
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiScaleDiscriminator(torch.nn.Module):
+
+    def __init__(self):
+        super(MultiScaleDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList([
+            DiscriminatorS(use_spectral_norm=True),
+            DiscriminatorS(),
+            DiscriminatorS(),
+        ])
+        from pytorch_wavelets import DWT1DForward
+        self.meanpools = nn.ModuleList(
+            [DWT1DForward(wave='db3', J=1),
+             DWT1DForward(wave='db3', J=1)])
+        self.convs = nn.ModuleList([
+            weight_norm(Conv1d(2, 1, 15, 1, padding=7)),
+            weight_norm(Conv1d(2, 1, 15, 1, padding=7))
+        ])
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            if i != 0:
+                yl, yh = self.meanpools[i - 1](y)
+                y = torch.cat([yl, yh[0]], dim=1)
+                y = self.convs[i - 1](y)
+                y = F.leaky_relu(y, LRELU_SLOPE)
+
+                yl_hat, yh_hat = self.meanpools[i - 1](y_hat)
+                y_hat = torch.cat([yl_hat, yh_hat[0]], dim=1)
+                y_hat = self.convs[i - 1](y_hat)
+                y_hat = F.leaky_relu(y_hat, LRELU_SLOPE)
+
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorSTFT(torch.nn.Module):
+
+    def __init__(self,
+                 kernel_size=11,
+                 stride=2,
+                 use_spectral_norm=False,
+                 fft_size=1024,
+                 shift_size=120,
+                 win_length=600,
+                 window='hann_window'):
+        super(DiscriminatorSTFT, self).__init__()
+        self.fft_size = fft_size
+        self.shift_size = shift_size
+        self.win_length = win_length
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(
+                Conv2d(
+                    fft_size // 2 + 1,
+                    32, (15, 1), (1, 1),
+                    padding=(get_padding(15, 1), 0))),
+            norm_f(
+                Conv2d(
+                    32,
+                    32, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(9, 1), 0))),
+            norm_f(
+                Conv2d(
+                    32,
+                    32, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(9, 1), 0))),
+            norm_f(
+                Conv2d(
+                    32,
+                    32, (kernel_size, 1), (stride, 1),
+                    padding=(get_padding(9, 1), 0))),
+            norm_f(Conv2d(32, 32, (5, 1), (1, 1), padding=(2, 0))),
+        ])
+        self.conv_post = norm_f(Conv2d(32, 1, (3, 1), (1, 1), padding=(1, 0)))
+        self.register_buffer('window', getattr(torch, window)(win_length))
+
+    def forward(self, wav):
+        wav = torch.squeeze(wav, 1)
+        x_mag = stft(wav, self.fft_size, self.shift_size, self.win_length,
+                     self.window)
+        x = torch.transpose(x_mag, 2, 1).unsqueeze(-1)
+        fmap = []
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = x.squeeze(-1)
+
+        return x, fmap
+
+
+class MultiSTFTDiscriminator(torch.nn.Module):
+
+    def __init__(
+        self,
+        fft_sizes=[1024, 2048, 512],
+        hop_sizes=[120, 240, 50],
+        win_lengths=[600, 1200, 240],
+        window='hann_window',
+    ):
+        super(MultiSTFTDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList()
+        for fs, ss, wl in zip(fft_sizes, hop_sizes, win_lengths):
+            self.discriminators += [
+                DiscriminatorSTFT(fft_size=fs, shift_size=ss, win_length=wl)
+            ]
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+def feature_loss(fmap_r, fmap_g):
+    loss = 0
+    for dr, dg in zip(fmap_r, fmap_g):
+        for rl, gl in zip(dr, dg):
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss * 2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+    loss = 0
+    r_losses = []
+    g_losses = []
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        r_loss = torch.mean((1 - dr)**2)
+        g_loss = torch.mean(dg**2)
+        loss += (r_loss + g_loss)
+        r_losses.append(r_loss.item())
+        g_losses.append(g_loss.item())
+
+    return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+    loss = 0
+    gen_losses = []
+    for dg in disc_outputs:
+        temp_loss = torch.mean((1 - dg)**2)
+        gen_losses.append(temp_loss)
+        loss += temp_loss
+
+    return loss, gen_losses

modelscope/models/audio/tts/vocoder/models/utils.py

@@ -0,0 +1,59 @@
+import glob
+import os
+
+import matplotlib
+import matplotlib.pylab as plt
+import torch
+from torch.nn.utils import weight_norm
+
+matplotlib.use('Agg')
+
+
+def plot_spectrogram(spectrogram):
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(
+        spectrogram, aspect='auto', origin='lower', interpolation='none')
+    plt.colorbar(im, ax=ax)
+
+    fig.canvas.draw()
+    plt.close()
+
+    return fig
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find('Conv') != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find('Conv') != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print('Complete.')
+    return checkpoint_dict
+
+
+def save_checkpoint(filepath, obj):
+    print('Saving checkpoint to {}'.format(filepath))
+    torch.save(obj, filepath)
+    print('Complete.')
+
+
+def scan_checkpoint(cp_dir, prefix):
+    pattern = os.path.join(cp_dir, prefix + '????????')
+    cp_list = glob.glob(pattern)
+    if len(cp_list) == 0:
+        return None
+    return sorted(cp_list)[-1]

modelscope/models/base.py

@@ -62,4 +62,6 @@ class Model(ABC):
         if hasattr(model_cfg, 'model_type') and not hasattr(model_cfg, 'type'):
             model_cfg.type = model_cfg.model_type
         model_cfg.model_dir = local_model_dir
+        for k, v in kwargs.items():
+            model_cfg.k = v
         return build_model(model_cfg, task_name)

modelscope/models/multi_model/__init__.py

@@ -0,0 +1 @@
+from .image_captioning_model import OfaForImageCaptioning

modelscope/models/multi_model/image_captioning_model.py

@@ -0,0 +1,80 @@
+import os.path as osp
+from typing import Any, Dict
+
+from PIL import Image
+
+from modelscope.utils.constant import ModelFile, Tasks
+from ..base import Model
+from ..builder import MODELS
+
+__all__ = ['OfaForImageCaptioning']
+
+
+@MODELS.register_module(
+    Tasks.image_captioning, module_name=r'ofa-image-captioning')
+class OfaForImageCaptioning(Model):
+
+    def __init__(self, model_dir, *args, **kwargs):
+        super().__init__(model_dir=model_dir, *args, **kwargs)
+        ckpt_name = ModelFile.TORCH_MODEL_FILE
+        local_model = osp.join(model_dir, ckpt_name)
+        bpe_dir = model_dir
+        # turn on cuda if GPU is available
+        from fairseq import checkpoint_utils, tasks, utils
+        from ofa.tasks.mm_tasks import CaptionTask
+        from ofa.utils.eval_utils import eval_caption
+        self.eval_caption = eval_caption
+
+        tasks.register_task('caption', CaptionTask)
+        use_cuda = kwargs['use_cuda'] if 'use_cuda' in kwargs else False
+        use_fp16 = kwargs[
+            'use_fp16'] if 'use_fp16' in kwargs and use_cuda else False
+        overrides = {
+            'bpe_dir': bpe_dir,
+            'eval_cider': False,
+            'beam': 5,
+            'max_len_b': 16,
+            'no_repeat_ngram_size': 3,
+            'seed': 7
+        }
+        models, cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+            utils.split_paths(local_model), arg_overrides=overrides)
+
+        # Move models to GPU
+        for model in models:
+            model.eval()
+            if use_cuda:
+                model.cuda()
+            if use_fp16:
+                model.half()
+            model.prepare_for_inference_(cfg)
+        self.models = models
+        # Initialize generator
+        self.generator = task.build_generator(models, cfg.generation)
+
+        # Initialize transform
+        from torchvision import transforms
+        mean = [0.5, 0.5, 0.5]
+        std = [0.5, 0.5, 0.5]
+
+        self.patch_resize_transform = transforms.Compose([
+            lambda image: image.convert('RGB'),
+            transforms.Resize(
+                (cfg.task.patch_image_size, cfg.task.patch_image_size),
+                interpolation=Image.BICUBIC),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=mean, std=std),
+        ])
+        self.task = task
+
+    def forward(self, input: Dict[str, Any]) -> Dict[str, Any]:
+        results, _ = self.eval_caption(self.task, self.generator, self.models,
+                                       input)
+        return {
+            'image_id': results[0]['image_id'],
+            'caption': results[0]['caption']
+        }
+
+    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        # What should we do here ?
+        return inputs

modelscope/models/nlp/__init__.py

@@ -1,4 +1,5 @@
-from .sentence_similarity_model import *  # noqa F403
-from .sequence_classification_model import *  # noqa F403
-from .text_generation_model import *  # noqa F403
+from .bert_for_sequence_classification import *  # noqa F403
+from .palm_for_text_generation import *  # noqa F403
+from .sbert_for_sentence_similarity import *  # noqa F403
+from .sbert_for_token_classification import *  # noqa F403
 from .zero_shot_classification_model import *  # noqa F403

modelscope/models/nlp/palm_for_text_generation.py

@@ -0,0 +1,43 @@
+from typing import Dict
+
+from modelscope.utils.constant import Tasks
+from ..base import Model, Tensor
+from ..builder import MODELS
+
+__all__ = ['PalmForTextGeneration']
+
+
+@MODELS.register_module(Tasks.text_generation, module_name=r'palm2.0')
+class PalmForTextGeneration(Model):
+
+    def __init__(self, model_dir: str, *args, **kwargs):
+        """initialize the text generation model from the `model_dir` path.
+
+        Args:
+            model_dir (str): the model path.
+            model_cls (Optional[Any], optional): model loader, if None, use the
+                default loader to load model weights, by default None.
+        """
+        super().__init__(model_dir, *args, **kwargs)
+        self.model_dir = model_dir
+
+        from sofa.models.palm_v2 import PalmForConditionalGeneration, Translator
+        model = PalmForConditionalGeneration.from_pretrained(model_dir)
+        self.tokenizer = model.tokenizer
+        self.generator = Translator(model)
+
+    def forward(self, input: Dict[str, Tensor]) -> Dict[str, Tensor]:
+        """return the result by the model
+
+        Args:
+            input (Dict[str, Tensor]): the preprocessed data
+
+        Returns:
+            Dict[str, Tensor]: results
+                Example:
+                    {
+                        'predictions': Tensor([[1377, 4959, 2785, 6392...])]), # tokens need to be decode by tokenizer
+                    }
+        """
+
+        return self.generator(**input)

modelscope/models/nlp/sbert_for_token_classification.py

@@ -0,0 +1,56 @@
+from typing import Any, Dict, Union
+
+import numpy as np
+import torch
+from sofa import SbertConfig, SbertForTokenClassification
+
+from modelscope.utils.constant import Tasks
+from ..base import Model, Tensor
+from ..builder import MODELS
+
+__all__ = ['StructBertForTokenClassification']
+
+
+@MODELS.register_module(
+    Tasks.word_segmentation,
+    module_name=r'structbert-chinese-word-segmentation')
+class StructBertForTokenClassification(Model):
+
+    def __init__(self, model_dir: str, *args, **kwargs):
+        """initialize the word segmentation model from the `model_dir` path.
+
+        Args:
+            model_dir (str): the model path.
+            model_cls (Optional[Any], optional): model loader, if None, use the
+                default loader to load model weights, by default None.
+        """
+        super().__init__(model_dir, *args, **kwargs)
+        self.model_dir = model_dir
+        self.model = SbertForTokenClassification.from_pretrained(
+            self.model_dir)
+        self.config = SbertConfig.from_pretrained(self.model_dir)
+
+    def forward(self, input: Dict[str,
+                                  Any]) -> Dict[str, Union[str, np.ndarray]]:
+        """return the result by the model
+
+        Args:
+            input (Dict[str, Any]): the preprocessed data
+
+        Returns:
+            Dict[str, Union[str,np.ndarray]]: results
+                Example:
+                    {
+                        'predictions': array([1,4]), # lable 0-negative 1-positive
+                        'logits': array([[-0.53860897,  1.5029076 ]], dtype=float32) # true value
+                        'text': str(今天),
+                    }
+        """
+        input_ids = torch.tensor(input['input_ids']).unsqueeze(0)
+        output = self.model(input_ids)
+        logits = output.logits
+        pred = torch.argmax(logits[0], dim=-1)
+        pred = pred.numpy()
+
+        rst = {'predictions': pred, 'logits': logits, 'text': input['text']}
+        return rst

modelscope/models/nlp/text_generation_model.py

@@ -1,52 +0,0 @@
-from typing import Any, Dict
-
-from modelscope.utils.constant import Tasks
-from ..base import Model, Tensor
-from ..builder import MODELS
-
-__all__ = ['PalmForTextGenerationModel']
-
-
-@MODELS.register_module(Tasks.text_generation, module_name=r'palm')
-class PalmForTextGenerationModel(Model):
-
-    def __init__(self, model_dir: str, *args, **kwargs):
-        """initialize the text generation model from the `model_dir` path.
-
-        Args:
-            model_dir (str): the model path.
-            model_cls (Optional[Any], optional): model loader, if None, use the
-                default loader to load model weights, by default None.
-        """
-        from sofa import PalmTokenizer
-
-        super().__init__(model_dir, *args, **kwargs)
-        self.model_dir = model_dir
-
-        from sofa.models.palm import PalmForConditionalGeneration, TextGenerator
-        tokenizer = kwargs.pop('tokenizer',
-                               PalmTokenizer.from_pretrained(model_dir))
-        model = PalmForConditionalGeneration.from_pretrained(model_dir)
-        self.generator = TextGenerator(model, tokenizer)
-
-    def forward(self, input: Dict[str, Tensor]) -> Dict[str, Tensor]:
-        """return the result by the model
-
-        Args:
-            input (Dict[str, Any]): the preprocessed data
-
-        Returns:
-            Dict[str, np.ndarray]: results
-                Example:
-                    {
-                        'predictions': array([1]), # lable 0-negative 1-positive
-                        'probabilities': array([[0.11491239, 0.8850876 ]], dtype=float32),
-                        'logits': array([[-0.53860897,  1.5029076 ]], dtype=float32) # true value
-                    }
-        """
-
-        encoder_inputs = [
-            input['input_ids'], input['token_type_ids'],
-            input['attention_mask']
-        ]
-        return self.generator(encoder_inputs)

modelscope/pipelines/__init__.py

@@ -1,4 +1,4 @@
-from .audio import *  # noqa F403
+from .audio import LinearAECPipeline
 from .base import Pipeline
 from .builder import pipeline
 from .cv import *  # noqa F403

modelscope/pipelines/audio/__init__.py

@@ -0,0 +1,2 @@
+from .linear_aec_pipeline import LinearAECPipeline
+from .text_to_speech_pipeline import *  # noqa F403

modelscope/pipelines/audio/linear_aec_pipeline.py

@@ -0,0 +1,160 @@
+import importlib
+import os
+from typing import Any, Dict
+
+import numpy as np
+import scipy.io.wavfile as wav
+import torch
+import yaml
+
+from modelscope.preprocessors.audio import LinearAECAndFbank
+from modelscope.utils.constant import ModelFile, Tasks
+from ..base import Pipeline
+from ..builder import PIPELINES
+
+FEATURE_MVN = 'feature.DEY.mvn.txt'
+
+CONFIG_YAML = 'dey_mini.yaml'
+
+
+def initialize_config(module_cfg):
+    r"""According to config items, load specific module dynamically with params.
+        1. Load the module corresponding to the "module" param.
+        2. Call function (or instantiate class) corresponding to the "main" param.
+        3. Send the param (in "args") into the function (or class) when calling ( or instantiating).
+
+    Args:
+        module_cfg (dict): config items, eg:
+            {
+                "module": "models.model",
+                "main": "Model",
+                "args": {...}
+            }
+
+    Returns:
+        the module loaded.
+    """
+    module = importlib.import_module(module_cfg['module'])
+    return getattr(module, module_cfg['main'])(**module_cfg['args'])
+
+
+@PIPELINES.register_module(
+    Tasks.speech_signal_process, module_name=r'speech_dfsmn_aec_psm_16k')
+class LinearAECPipeline(Pipeline):
+    r"""AEC Inference Pipeline only support 16000 sample rate.
+
+    When invoke the class with pipeline.__call__(), you should provide two params:
+        Dict[str, Any]
+            the path of wav files，eg:{
+            "nearend_mic": "/your/data/near_end_mic_audio.wav",
+            "farend_speech": "/your/data/far_end_speech_audio.wav"}
+        output_path (str, optional): "/your/output/audio_after_aec.wav"
+            the file path to write generate audio.
+    """
+
+    def __init__(self, model):
+        r"""
+        Args:
+            model: model id on modelscope hub.
+        """
+        super().__init__(model=model)
+        self.use_cuda = torch.cuda.is_available()
+        with open(
+                os.path.join(self.model, CONFIG_YAML), encoding='utf-8') as f:
+            self.config = yaml.full_load(f.read())
+            self.config['io']['mvn'] = os.path.join(self.model, FEATURE_MVN)
+        self._init_model()
+        self.preprocessor = LinearAECAndFbank(self.config['io'])
+
+        n_fft = self.config['loss']['args']['n_fft']
+        hop_length = self.config['loss']['args']['hop_length']
+        winlen = n_fft
+        window = torch.hamming_window(winlen, periodic=False)
+
+        def stft(x):
+            return torch.stft(
+                x,
+                n_fft,
+                hop_length,
+                winlen,
+                center=False,
+                window=window.to(x.device),
+                return_complex=False)
+
+        def istft(x, slen):
+            return torch.istft(
+                x,
+                n_fft,
+                hop_length,
+                winlen,
+                window=window.to(x.device),
+                center=False,
+                length=slen)
+
+        self.stft = stft
+        self.istft = istft
+
+    def _init_model(self):
+        checkpoint = torch.load(
+            os.path.join(self.model, ModelFile.TORCH_MODEL_BIN_FILE),
+            map_location='cpu')
+        self.model = initialize_config(self.config['nnet'])
+        if self.use_cuda:
+            self.model = self.model.cuda()
+        self.model.load_state_dict(checkpoint)
+
+    def forward(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        r"""The AEC process.
+
+        Args:
+            inputs: dict={'feature': Tensor, 'base': Tensor}
+                'feature' feature of input audio.
+                'base' the base audio to mask.
+
+        Returns:
+            dict:
+                {
+                    'output_pcm': generated audio array
+                }
+        """
+        output_data = self._process(inputs['feature'], inputs['base'])
+        return {'output_pcm': output_data}
+
+    def postprocess(self, inputs: Dict[str, Any], **kwargs) -> Dict[str, Any]:
+        r"""The post process. Will save audio to file, if the output_path is given.
+
+        Args:
+            inputs: dict:
+                {
+                    'output_pcm': generated audio array
+                }
+            kwargs: accept 'output_path' which is the path to write generated audio
+
+        Returns:
+            dict:
+                {
+                    'output_pcm': generated audio array
+                }
+        """
+        if 'output_path' in kwargs.keys():
+            wav.write(kwargs['output_path'], self.preprocessor.SAMPLE_RATE,
+                      inputs['output_pcm'].astype(np.int16))
+        inputs['output_pcm'] = inputs['output_pcm'] / 32768.0
+        return inputs
+
+    def _process(self, fbanks, mixture):
+        if self.use_cuda:
+            fbanks = fbanks.cuda()
+            mixture = mixture.cuda()
+        if self.model.vad:
+            with torch.no_grad():
+                masks, vad = self.model(fbanks.unsqueeze(0))
+                masks = masks.permute([2, 1, 0])
+        else:
+            with torch.no_grad():
+                masks = self.model(fbanks.unsqueeze(0))
+                masks = masks.permute([2, 1, 0])
+        spectrum = self.stft(mixture)
+        masked_spec = spectrum * masks
+        masked_sig = self.istft(masked_spec, len(mixture)).cpu().numpy()
+        return masked_sig

modelscope/pipelines/audio/text_to_speech_pipeline.py

@@ -0,0 +1,46 @@
+import time
+from typing import Any, Dict, List
+
+import numpy as np
+
+from modelscope.models import Model
+from modelscope.models.audio.tts.am import SambertNetHifi16k
+from modelscope.models.audio.tts.vocoder import Hifigan16k
+from modelscope.pipelines.base import Pipeline
+from modelscope.pipelines.builder import PIPELINES
+from modelscope.preprocessors import TextToTacotronSymbols, build_preprocessor
+from modelscope.utils.constant import Fields, Tasks
+
+__all__ = ['TextToSpeechSambertHifigan16kPipeline']
+
+
+@PIPELINES.register_module(
+    Tasks.text_to_speech, module_name=r'tts-sambert-hifigan-16k')
+class TextToSpeechSambertHifigan16kPipeline(Pipeline):
+
+    def __init__(self,
+                 config_file: str = None,
+                 model: List[Model] = None,
+                 preprocessor: TextToTacotronSymbols = None,
+                 **kwargs):
+        super().__init__(
+            config_file=config_file,
+            model=model,
+            preprocessor=preprocessor,
+            **kwargs)
+        assert len(model) == 2, 'model number should be 2'
+        self._am = model[0]
+        self._vocoder = model[1]
+        self._preprocessor = preprocessor
+
+    def forward(self, inputs: Dict[str, Any]) -> Dict[str, np.ndarray]:
+        texts = inputs['texts']
+        audio_total = np.empty((0), dtype='int16')
+        for line in texts:
+            line = line.strip().split('\t')
+            audio = self._vocoder.forward(self._am.forward(line[1]))
+            audio_total = np.append(audio_total, audio, axis=0)
+        return {'output': audio_total}
+
+    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        return inputs

modelscope/pipelines/builder.py

@@ -13,20 +13,26 @@ PIPELINES = Registry('pipelines')
 
 DEFAULT_MODEL_FOR_PIPELINE = {
     # TaskName: (pipeline_module_name, model_repo)
+    Tasks.word_segmentation:
+    ('structbert-chinese-word-segmentation',
+     'damo/nlp_structbert_word-segmentation_chinese-base'),
     Tasks.sentence_similarity:
     ('sbert-base-chinese-sentence-similarity',
      'damo/nlp_structbert_sentence-similarity_chinese-base'),
-    Tasks.image_matting: ('image-matting', 'damo/cv_unet_image-matting_damo'),
-    Tasks.text_classification:
-    ('bert-sentiment-analysis', 'damo/bert-base-sst2'),
+    Tasks.image_matting: ('image-matting', 'damo/cv_unet_image-matting'),
+    Tasks.text_classification: ('bert-sentiment-analysis',
+                                'damo/bert-base-sst2'),
     Tasks.zero_shot_classification:
     ('bert-zero-shot-classification',
      'damo/nlp_structbert_zero-shot-classification_chinese-base'),
-    Tasks.text_generation: ('palm', 'damo/nlp_palm_text-generation_chinese'),
-    Tasks.image_captioning: ('ofa', None),
+    Tasks.text_generation: ('palm2.0',
+                            'damo/nlp_palm2.0_text-generation_chinese-base'),
+    Tasks.image_captioning: ('ofa', 'damo/ofa_image-caption_coco_large_en'),
     Tasks.image_generation:
     ('person-image-cartoon',
      'damo/cv_unet_person-image-cartoon_compound-models'),
+    Tasks.ocr_detection: ('ocr-detection',
+                          'damo/cv_resnet18_ocr-detection-line-level_damo'),
 }
 
 

modelscope/pipelines/cv/__init__.py

@@ -1,2 +1,3 @@
 from .image_cartoon_pipeline import ImageCartoonPipeline
 from .image_matting_pipeline import ImageMattingPipeline
+from .ocr_detection_pipeline import OCRDetectionPipeline

modelscope/pipelines/cv/ocr_detection_pipeline.py

@@ -0,0 +1,167 @@
+import math
+import os
+import os.path as osp
+import sys
+from typing import Any, Dict, List, Tuple, Union
+
+import cv2
+import numpy as np
+import PIL
+import tensorflow as tf
+import tf_slim as slim
+
+from modelscope.pipelines.base import Input
+from modelscope.preprocessors import load_image
+from modelscope.utils.constant import ModelFile, Tasks
+from modelscope.utils.logger import get_logger
+from ..base import Pipeline
+from ..builder import PIPELINES
+from .ocr_utils import model_resnet_mutex_v4_linewithchar, ops, utils
+
+if tf.__version__ >= '2.0':
+    tf = tf.compat.v1
+tf.compat.v1.disable_eager_execution()
+
+logger = get_logger()
+
+# constant
+RBOX_DIM = 5
+OFFSET_DIM = 6
+WORD_POLYGON_DIM = 8
+OFFSET_VARIANCE = [0.1, 0.1, 0.1, 0.1, 0.1, 0.1]
+
+FLAGS = tf.app.flags.FLAGS
+tf.app.flags.DEFINE_float('node_threshold', 0.4,
+                          'Confidence threshold for nodes')
+tf.app.flags.DEFINE_float('link_threshold', 0.6,
+                          'Confidence threshold for links')
+
+
+@PIPELINES.register_module(
+    Tasks.ocr_detection, module_name=Tasks.ocr_detection)
+class OCRDetectionPipeline(Pipeline):
+
+    def __init__(self, model: str):
+        super().__init__(model=model)
+        model_path = osp.join(
+            osp.join(self.model, ModelFile.TF_CHECKPOINT_FOLDER),
+            'checkpoint-80000')
+
+        config = tf.ConfigProto(allow_soft_placement=True)
+        config.gpu_options.allow_growth = True
+        self._session = tf.Session(config=config)
+        global_step = tf.get_variable(
+            'global_step', [],
+            initializer=tf.constant_initializer(0),
+            dtype=tf.int64,
+            trainable=False)
+        variable_averages = tf.train.ExponentialMovingAverage(
+            0.997, global_step)
+        self.input_images = tf.placeholder(
+            tf.float32, shape=[1, 1024, 1024, 3], name='input_images')
+        self.output = {}
+
+        # detector
+        detector = model_resnet_mutex_v4_linewithchar.SegLinkDetector()
+        all_maps = detector.build_model(self.input_images, is_training=False)
+
+        # decode local predictions
+        all_nodes, all_links, all_reg = [], [], []
+        for i, maps in enumerate(all_maps):
+            cls_maps, lnk_maps, reg_maps = maps[0], maps[1], maps[2]
+            reg_maps = tf.multiply(reg_maps, OFFSET_VARIANCE)
+
+            cls_prob = tf.nn.softmax(tf.reshape(cls_maps, [-1, 2]))
+
+            lnk_prob_pos = tf.nn.softmax(tf.reshape(lnk_maps, [-1, 4])[:, :2])
+            lnk_prob_mut = tf.nn.softmax(tf.reshape(lnk_maps, [-1, 4])[:, 2:])
+            lnk_prob = tf.concat([lnk_prob_pos, lnk_prob_mut], axis=1)
+
+            all_nodes.append(cls_prob)
+            all_links.append(lnk_prob)
+            all_reg.append(reg_maps)
+
+        # decode segments and links
+        image_size = tf.shape(self.input_images)[1:3]
+        segments, group_indices, segment_counts, _ = ops.decode_segments_links_python(
+            image_size,
+            all_nodes,
+            all_links,
+            all_reg,
+            anchor_sizes=list(detector.anchor_sizes))
+
+        # combine segments
+        combined_rboxes, combined_counts = ops.combine_segments_python(
+            segments, group_indices, segment_counts)
+        self.output['combined_rboxes'] = combined_rboxes
+        self.output['combined_counts'] = combined_counts
+
+        with self._session.as_default() as sess:
+            logger.info(f'loading model from {model_path}')
+            # load model
+            model_loader = tf.train.Saver(
+                variable_averages.variables_to_restore())
+            model_loader.restore(sess, model_path)
+
+    def preprocess(self, input: Input) -> Dict[str, Any]:
+        if isinstance(input, str):
+            img = np.array(load_image(input))
+        elif isinstance(input, PIL.Image.Image):
+            img = np.array(input.convert('RGB'))
+        elif isinstance(input, np.ndarray):
+            if len(input.shape) == 2:
+                img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+            img = input[:, :, ::-1]  # in rgb order
+        else:
+            raise TypeError(f'input should be either str, PIL.Image,'
+                            f' np.array, but got {type(input)}')
+        h, w, c = img.shape
+        img_pad = np.zeros((max(h, w), max(h, w), 3), dtype=np.float32)
+        img_pad[:h, :w, :] = img
+
+        resize_size = 1024
+        img_pad_resize = cv2.resize(img_pad, (resize_size, resize_size))
+        img_pad_resize = cv2.cvtColor(img_pad_resize, cv2.COLOR_RGB2BGR)
+        img_pad_resize = img_pad_resize - np.array([123.68, 116.78, 103.94],
+                                                   dtype=np.float32)
+
+        resize_size = tf.stack([resize_size, resize_size])
+        orig_size = tf.stack([max(h, w), max(h, w)])
+        self.output['orig_size'] = orig_size
+        self.output['resize_size'] = resize_size
+
+        result = {'img': np.expand_dims(img_pad_resize, axis=0)}
+        return result
+
+    def forward(self, input: Dict[str, Any]) -> Dict[str, Any]:
+        with self._session.as_default():
+            feed_dict = {self.input_images: input['img']}
+            sess_outputs = self._session.run(self.output, feed_dict=feed_dict)
+            return sess_outputs
+
+    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        rboxes = inputs['combined_rboxes'][0]
+        count = inputs['combined_counts'][0]
+        rboxes = rboxes[:count, :]
+
+        # convert rboxes to polygons and find its coordinates on the original image
+        orig_h, orig_w = inputs['orig_size']
+        resize_h, resize_w = inputs['resize_size']
+        polygons = utils.rboxes_to_polygons(rboxes)
+        scale_y = float(orig_h) / float(resize_h)
+        scale_x = float(orig_w) / float(resize_w)
+
+        # confine polygons inside image
+        polygons[:, ::2] = np.maximum(
+            0, np.minimum(polygons[:, ::2] * scale_x, orig_w - 1))
+        polygons[:, 1::2] = np.maximum(
+            0, np.minimum(polygons[:, 1::2] * scale_y, orig_h - 1))
+        polygons = np.round(polygons).astype(np.int32)
+
+        # nms
+        dt_n9 = [o + [utils.cal_width(o)] for o in polygons.tolist()]
+        dt_nms = utils.nms_python(dt_n9)
+        dt_polygons = np.array([o[:8] for o in dt_nms])
+
+        result = {'det_polygons': dt_polygons}
+        return result

modelscope/pipelines/cv/ocr_utils/model_resnet_mutex_v4_linewithchar.py

@@ -0,0 +1,158 @@
+import tensorflow as tf
+import tf_slim as slim
+
+from . import ops, resnet18_v1, resnet_utils
+
+if tf.__version__ >= '2.0':
+    tf = tf.compat.v1
+
+# constants
+OFFSET_DIM = 6
+
+N_LOCAL_LINKS = 8
+N_CROSS_LINKS = 4
+N_SEG_CLASSES = 2
+N_LNK_CLASSES = 4
+
+POS_LABEL = 1
+NEG_LABEL = 0
+
+
+class SegLinkDetector():
+
+    def __init__(self):
+        self.anchor_sizes = [6., 11.84210526, 23.68421053, 45., 90., 150.]
+
+    def _detection_classifier(self,
+                              maps,
+                              ksize,
+                              weight_decay,
+                              cross_links=False,
+                              scope=None):
+
+        with tf.variable_scope(scope):
+            seg_depth = N_SEG_CLASSES
+            if cross_links:
+                lnk_depth = N_LNK_CLASSES * (N_LOCAL_LINKS + N_CROSS_LINKS)
+            else:
+                lnk_depth = N_LNK_CLASSES * N_LOCAL_LINKS
+            reg_depth = OFFSET_DIM
+            map_depth = maps.get_shape()[3]
+            inter_maps, inter_relu = ops.conv2d(
+                maps, map_depth, 256, 1, 1, 'SAME', scope='conv_inter')
+
+            dir_maps, dir_relu = ops.conv2d(
+                inter_relu, 256, 2, ksize, 1, 'SAME', scope='conv_dir')
+            cen_maps, cen_relu = ops.conv2d(
+                inter_relu, 256, 2, ksize, 1, 'SAME', scope='conv_cen')
+            pol_maps, pol_relu = ops.conv2d(
+                inter_relu, 256, 8, ksize, 1, 'SAME', scope='conv_pol')
+            concat_relu = tf.concat([dir_relu, cen_relu, pol_relu], axis=-1)
+            _, lnk_embedding = ops.conv_relu(
+                concat_relu, 12, 256, 1, 1, scope='lnk_embedding')
+            lnk_maps, lnk_relu = ops.conv2d(
+                inter_relu + lnk_embedding,
+                256,
+                lnk_depth,
+                ksize,
+                1,
+                'SAME',
+                scope='conv_lnk')
+
+            char_seg_maps, char_seg_relu = ops.conv2d(
+                inter_relu,
+                256,
+                seg_depth,
+                ksize,
+                1,
+                'SAME',
+                scope='conv_char_cls')
+            char_reg_maps, char_reg_relu = ops.conv2d(
+                inter_relu,
+                256,
+                reg_depth,
+                ksize,
+                1,
+                'SAME',
+                scope='conv_char_reg')
+            concat_char_relu = tf.concat([char_seg_relu, char_reg_relu],
+                                         axis=-1)
+            _, char_embedding = ops.conv_relu(
+                concat_char_relu, 8, 256, 1, 1, scope='conv_char_embedding')
+            seg_maps, seg_relu = ops.conv2d(
+                inter_relu + char_embedding,
+                256,
+                seg_depth,
+                ksize,
+                1,
+                'SAME',
+                scope='conv_cls')
+            reg_maps, reg_relu = ops.conv2d(
+                inter_relu + char_embedding,
+                256,
+                reg_depth,
+                ksize,
+                1,
+                'SAME',
+                scope='conv_reg')
+
+        return seg_relu, lnk_relu, reg_relu
+
+    def _build_cnn(self, images, weight_decay, is_training):
+        with slim.arg_scope(
+                resnet18_v1.resnet_arg_scope(weight_decay=weight_decay)):
+            logits, end_points = resnet18_v1.resnet_v1_18(
+                images, is_training=is_training, scope='resnet_v1_18')
+
+        outputs = {
+            'conv3_3': end_points['pool1'],
+            'conv4_3': end_points['pool2'],
+            'fc7': end_points['pool3'],
+            'conv8_2': end_points['pool4'],
+            'conv9_2': end_points['pool5'],
+            'conv10_2': end_points['pool6'],
+        }
+        return outputs
+
+    def build_model(self, images, is_training=True, scope=None):
+
+        weight_decay = 5e-4  # FLAGS.weight_decay
+        cnn_outputs = self._build_cnn(images, weight_decay, is_training)
+        det_0 = self._detection_classifier(
+            cnn_outputs['conv3_3'],
+            3,
+            weight_decay,
+            cross_links=False,
+            scope='dete_0')
+        det_1 = self._detection_classifier(
+            cnn_outputs['conv4_3'],
+            3,
+            weight_decay,
+            cross_links=True,
+            scope='dete_1')
+        det_2 = self._detection_classifier(
+            cnn_outputs['fc7'],
+            3,
+            weight_decay,
+            cross_links=True,
+            scope='dete_2')
+        det_3 = self._detection_classifier(
+            cnn_outputs['conv8_2'],
+            3,
+            weight_decay,
+            cross_links=True,
+            scope='dete_3')
+        det_4 = self._detection_classifier(
+            cnn_outputs['conv9_2'],
+            3,
+            weight_decay,
+            cross_links=True,
+            scope='dete_4')
+        det_5 = self._detection_classifier(
+            cnn_outputs['conv10_2'],
+            3,
+            weight_decay,
+            cross_links=True,
+            scope='dete_5')
+        outputs = [det_0, det_1, det_2, det_3, det_4, det_5]
+        return outputs

modelscope/pipelines/cv/ocr_utils/resnet18_v1.py

@@ -0,0 +1,432 @@
+"""Contains definitions for the original form of Residual Networks.
+The 'v1' residual networks (ResNets) implemented in this module were proposed
+by:
+[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
+    Deep Residual Learning for Image Recognition. arXiv:1512.03385
+Other variants were introduced in:
+[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
+    Identity Mappings in Deep Residual Networks. arXiv: 1603.05027
+The networks defined in this module utilize the bottleneck building block of
+[1] with projection shortcuts only for increasing depths. They employ batch
+normalization *after* every weight layer. This is the architecture used by
+MSRA in the Imagenet and MSCOCO 2016 competition models ResNet-101 and
+ResNet-152. See [2; Fig. 1a] for a comparison between the current 'v1'
+architecture and the alternative 'v2' architecture of [2] which uses batch
+normalization *before* every weight layer in the so-called full pre-activation
+units.
+Typical use:
+   from tensorflow.contrib.slim.nets import resnet_v1
+ResNet-101 for image classification into 1000 classes:
+   # inputs has shape [batch, 224, 224, 3]
+   with slim.arg_scope(resnet_v1.resnet_arg_scope()):
+      net, end_points = resnet_v1.resnet_v1_101(inputs, 1000, is_training=False)
+ResNet-101 for semantic segmentation into 21 classes:
+   # inputs has shape [batch, 513, 513, 3]
+   with slim.arg_scope(resnet_v1.resnet_arg_scope()):
+      net, end_points = resnet_v1.resnet_v1_101(inputs,
+                                                21,
+                                                is_training=False,
+                                                global_pool=False,
+                                                output_stride=16)
+"""
+import tensorflow as tf
+import tf_slim as slim
+
+from . import resnet_utils
+
+if tf.__version__ >= '2.0':
+    tf = tf.compat.v1
+
+resnet_arg_scope = resnet_utils.resnet_arg_scope
+
+
+@slim.add_arg_scope
+def basicblock(inputs,
+               depth,
+               depth_bottleneck,
+               stride,
+               rate=1,
+               outputs_collections=None,
+               scope=None):
+    """Bottleneck residual unit variant with BN after convolutions.
+    This is the original residual unit proposed in [1]. See Fig. 1(a) of [2] for
+    its definition. Note that we use here the bottleneck variant which has an
+    extra bottleneck layer.
+    When putting together two consecutive ResNet blocks that use this unit, one
+    should use stride = 2 in the last unit of the first block.
+    Args:
+      inputs: A tensor of size [batch, height, width, channels].
+      depth: The depth of the ResNet unit output.
+      depth_bottleneck: The depth of the bottleneck layers.
+      stride: The ResNet unit's stride. Determines the amount of downsampling of
+        the units output compared to its input.
+      rate: An integer, rate for atrous convolution.
+      outputs_collections: Collection to add the ResNet unit output.
+      scope: Optional variable_scope.
+    Returns:
+      The ResNet unit's output.
+    """
+    with tf.variable_scope(scope, 'bottleneck_v1', [inputs]) as sc:
+        depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
+        if depth == depth_in:
+            shortcut = resnet_utils.subsample(inputs, stride, 'shortcut')
+        else:
+            shortcut = slim.conv2d(
+                inputs,
+                depth, [1, 1],
+                stride=stride,
+                activation_fn=None,
+                scope='shortcut')
+
+        residual = resnet_utils.conv2d_same(
+            inputs, depth, 3, stride, rate=rate, scope='conv1')
+        residual = resnet_utils.conv2d_same(
+            residual, depth, 3, 1, rate=rate, scope='conv2')
+
+        output = tf.nn.relu(residual + shortcut)
+
+        return slim.utils.collect_named_outputs(outputs_collections,
+                                                sc.original_name_scope, output)
+
+
+@slim.add_arg_scope
+def bottleneck(inputs,
+               depth,
+               depth_bottleneck,
+               stride,
+               rate=1,
+               outputs_collections=None,
+               scope=None):
+    """Bottleneck residual unit variant with BN after convolutions.
+    This is the original residual unit proposed in [1]. See Fig. 1(a) of [2] for
+    its definition. Note that we use here the bottleneck variant which has an
+    extra bottleneck layer.
+    When putting together two consecutive ResNet blocks that use this unit, one
+    should use stride = 2 in the last unit of the first block.
+    Args:
+      inputs: A tensor of size [batch, height, width, channels].
+      depth: The depth of the ResNet unit output.
+      depth_bottleneck: The depth of the bottleneck layers.
+      stride: The ResNet unit's stride. Determines the amount of downsampling of
+        the units output compared to its input.
+      rate: An integer, rate for atrous convolution.
+      outputs_collections: Collection to add the ResNet unit output.
+      scope: Optional variable_scope.
+    Returns:
+      The ResNet unit's output.
+    """
+    with tf.variable_scope(scope, 'bottleneck_v1', [inputs]) as sc:
+        depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
+        if depth == depth_in:
+            shortcut = resnet_utils.subsample(inputs, stride, 'shortcut')
+        else:
+            shortcut = slim.conv2d(
+                inputs,
+                depth, [1, 1],
+                stride=stride,
+                activation_fn=None,
+                scope='shortcut')
+
+        residual = slim.conv2d(
+            inputs, depth_bottleneck, [1, 1], stride=1, scope='conv1')
+        residual = resnet_utils.conv2d_same(
+            residual, depth_bottleneck, 3, stride, rate=rate, scope='conv2')
+        residual = slim.conv2d(
+            residual,
+            depth, [1, 1],
+            stride=1,
+            activation_fn=None,
+            scope='conv3')
+
+        output = tf.nn.relu(shortcut + residual)
+
+        return slim.utils.collect_named_outputs(outputs_collections,
+                                                sc.original_name_scope, output)
+
+
+def resnet_v1(inputs,
+              blocks,
+              num_classes=None,
+              is_training=True,
+              global_pool=True,
+              output_stride=None,
+              include_root_block=True,
+              spatial_squeeze=True,
+              reuse=None,
+              scope=None):
+    """Generator for v1 ResNet models.
+    This function generates a family of ResNet v1 models. See the resnet_v1_*()
+    methods for specific model instantiations, obtained by selecting different
+    block instantiations that produce ResNets of various depths.
+    Training for image classification on Imagenet is usually done with [224, 224]
+    inputs, resulting in [7, 7] feature maps at the output of the last ResNet
+    block for the ResNets defined in [1] that have nominal stride equal to 32.
+    However, for dense prediction tasks we advise that one uses inputs with
+    spatial dimensions that are multiples of 32 plus 1, e.g., [321, 321]. In
+    this case the feature maps at the ResNet output will have spatial shape
+    [(height - 1) / output_stride + 1, (width - 1) / output_stride + 1]
+    and corners exactly aligned with the input image corners, which greatly
+    facilitates alignment of the features to the image. Using as input [225, 225]
+    images results in [8, 8] feature maps at the output of the last ResNet block.
+    For dense prediction tasks, the ResNet needs to run in fully-convolutional
+    (FCN) mode and global_pool needs to be set to False. The ResNets in [1, 2] all
+    have nominal stride equal to 32 and a good choice in FCN mode is to use
+    output_stride=16 in order to increase the density of the computed features at
+    small computational and memory overhead, cf. http://arxiv.org/abs/1606.00915.
+    Args:
+      inputs: A tensor of size [batch, height_in, width_in, channels].
+      blocks: A list of length equal to the number of ResNet blocks. Each element
+        is a resnet_utils.Block object describing the units in the block.
+      num_classes: Number of predicted classes for classification tasks. If None
+        we return the features before the logit layer.
+      is_training: whether is training or not.
+      global_pool: If True, we perform global average pooling before computing the
+        logits. Set to True for image classification, False for dense prediction.
+      output_stride: If None, then the output will be computed at the nominal
+        network stride. If output_stride is not None, it specifies the requested
+        ratio of input to output spatial resolution.
+      include_root_block: If True, include the initial convolution followed by
+        max-pooling, if False excludes it.
+      spatial_squeeze: if True, logits is of shape [B, C], if false logits is
+          of shape [B, 1, 1, C], where B is batch_size and C is number of classes.
+      reuse: whether or not the network and its variables should be reused. To be
+        able to reuse 'scope' must be given.
+      scope: Optional variable_scope.
+    Returns:
+      net: A rank-4 tensor of size [batch, height_out, width_out, channels_out].
+        If global_pool is False, then height_out and width_out are reduced by a
+        factor of output_stride compared to the respective height_in and width_in,
+        else both height_out and width_out equal one. If num_classes is None, then
+        net is the output of the last ResNet block, potentially after global
+        average pooling. If num_classes is not None, net contains the pre-softmax
+        activations.
+      end_points: A dictionary from components of the network to the corresponding
+        activation.
+    Raises:
+      ValueError: If the target output_stride is not valid.
+    """
+    with tf.variable_scope(scope, 'resnet_v1', [inputs], reuse=reuse) as sc:
+        end_points_collection = sc.name + '_end_points'
+        with slim.arg_scope(
+            [slim.conv2d, bottleneck, resnet_utils.stack_blocks_dense],
+                outputs_collections=end_points_collection):
+            with slim.arg_scope([slim.batch_norm], is_training=is_training):
+                net = inputs
+                if include_root_block:
+                    if output_stride is not None:
+                        if output_stride % 4 != 0:
+                            raise ValueError(
+                                'The output_stride needs to be a multiple of 4.'
+                            )
+                        output_stride /= 4
+                    net = resnet_utils.conv2d_same(
+                        net, 64, 7, stride=2, scope='conv1')
+                    net = tf.pad(net, [[0, 0], [1, 1], [1, 1], [0, 0]])
+                    net = slim.max_pool2d(net, [3, 3], stride=2, scope='pool1')
+
+                    net = slim.utils.collect_named_outputs(
+                        end_points_collection, 'pool2', net)
+
+                net = resnet_utils.stack_blocks_dense(net, blocks,
+                                                      output_stride)
+
+                end_points = slim.utils.convert_collection_to_dict(
+                    end_points_collection)
+
+                end_points['pool1'] = end_points['resnet_v1_18/block2/unit_2']
+                end_points['pool2'] = end_points['resnet_v1_18/block3/unit_2']
+                end_points['pool3'] = end_points['resnet_v1_18/block4/unit_2']
+                end_points['pool4'] = end_points['resnet_v1_18/block5/unit_2']
+                end_points['pool5'] = end_points['resnet_v1_18/block6/unit_2']
+                end_points['pool6'] = net
+
+                return net, end_points
+
+
+resnet_v1.default_image_size = 224
+
+
+def resnet_v1_18(inputs,
+                 num_classes=None,
+                 is_training=True,
+                 global_pool=True,
+                 output_stride=None,
+                 spatial_squeeze=True,
+                 reuse=None,
+                 scope='resnet_v1_18'):
+    """ResNet-18 model of [1]. See resnet_v1() for arg and return description."""
+    blocks = [
+        resnet_utils.Block('block1', basicblock,
+                           [(64, 64, 1)] + [(64, 64, 1)]),
+        resnet_utils.Block('block2', basicblock,
+                           [(128, 128, 1)] + [(128, 128, 1)]),
+        resnet_utils.Block('block3', basicblock,
+                           [(256, 256, 2)] + [(256, 256, 1)]),
+        resnet_utils.Block('block4', basicblock,
+                           [(512, 512, 2)] + [(512, 512, 1)]),
+        resnet_utils.Block('block5', basicblock,
+                           [(256, 256, 2)] + [(256, 256, 1)]),
+        resnet_utils.Block('block6', basicblock,
+                           [(256, 256, 2)] + [(256, 256, 1)]),
+        resnet_utils.Block('block7', basicblock,
+                           [(256, 256, 2)] + [(256, 256, 1)]),
+    ]
+    return resnet_v1(
+        inputs,
+        blocks,
+        num_classes,
+        is_training,
+        global_pool=global_pool,
+        output_stride=output_stride,
+        include_root_block=True,
+        spatial_squeeze=spatial_squeeze,
+        reuse=reuse,
+        scope=scope)
+
+
+resnet_v1_18.default_image_size = resnet_v1.default_image_size
+
+
+def resnet_v1_50(inputs,
+                 num_classes=None,
+                 is_training=True,
+                 global_pool=True,
+                 output_stride=None,
+                 spatial_squeeze=True,
+                 reuse=None,
+                 scope='resnet_v1_50'):
+    """ResNet-50 model of [1]. See resnet_v1() for arg and return description."""
+    blocks = [
+        resnet_utils.Block('block1', bottleneck,
+                           [(256, 64, 1)] * 2 + [(256, 64, 2)]),
+        resnet_utils.Block('block2', bottleneck,
+                           [(512, 128, 1)] * 3 + [(512, 128, 2)]),
+        resnet_utils.Block('block3', bottleneck,
+                           [(1024, 256, 1)] * 5 + [(1024, 256, 2)]),
+        resnet_utils.Block('block4', bottleneck,
+                           [(2048, 512, 1)] * 3 + [(2048, 512, 2)]),
+        resnet_utils.Block('block5', bottleneck,
+                           [(1024, 256, 1)] * 2 + [(1024, 256, 2)]),
+        resnet_utils.Block('block6', bottleneck, [(1024, 256, 1)] * 2),
+    ]
+    return resnet_v1(
+        inputs,
+        blocks,
+        num_classes,
+        is_training,
+        global_pool=global_pool,
+        output_stride=output_stride,
+        include_root_block=True,
+        spatial_squeeze=spatial_squeeze,
+        reuse=reuse,
+        scope=scope)
+
+
+resnet_v1_50.default_image_size = resnet_v1.default_image_size
+
+
+def resnet_v1_101(inputs,
+                  num_classes=None,
+                  is_training=True,
+                  global_pool=True,
+                  output_stride=None,
+                  spatial_squeeze=True,
+                  reuse=None,
+                  scope='resnet_v1_101'):
+    """ResNet-101 model of [1]. See resnet_v1() for arg and return description."""
+    blocks = [
+        resnet_utils.Block('block1', bottleneck,
+                           [(256, 64, 1)] * 2 + [(256, 64, 2)]),
+        resnet_utils.Block('block2', bottleneck,
+                           [(512, 128, 1)] * 3 + [(512, 128, 2)]),
+        resnet_utils.Block('block3', bottleneck,
+                           [(1024, 256, 1)] * 22 + [(1024, 256, 2)]),
+        resnet_utils.Block('block4', bottleneck, [(2048, 512, 1)] * 3)
+    ]
+    return resnet_v1(
+        inputs,
+        blocks,
+        num_classes,
+        is_training,
+        global_pool=global_pool,
+        output_stride=output_stride,
+        include_root_block=True,
+        spatial_squeeze=spatial_squeeze,
+        reuse=reuse,
+        scope=scope)
+
+
+resnet_v1_101.default_image_size = resnet_v1.default_image_size
+
+
+def resnet_v1_152(inputs,
+                  num_classes=None,
+                  is_training=True,
+                  global_pool=True,
+                  output_stride=None,
+                  spatial_squeeze=True,
+                  reuse=None,
+                  scope='resnet_v1_152'):
+    """ResNet-152 model of [1]. See resnet_v1() for arg and return description."""
+    blocks = [
+        resnet_utils.Block('block1', bottleneck,
+                           [(256, 64, 1)] * 2 + [(256, 64, 2)]),
+        resnet_utils.Block('block2', bottleneck,
+                           [(512, 128, 1)] * 7 + [(512, 128, 2)]),
+        resnet_utils.Block('block3', bottleneck,
+                           [(1024, 256, 1)] * 35 + [(1024, 256, 2)]),
+        resnet_utils.Block('block4', bottleneck, [(2048, 512, 1)] * 3)
+    ]
+    return resnet_v1(
+        inputs,
+        blocks,
+        num_classes,
+        is_training,
+        global_pool=global_pool,
+        output_stride=output_stride,
+        include_root_block=True,
+        spatial_squeeze=spatial_squeeze,
+        reuse=reuse,
+        scope=scope)
+
+
+resnet_v1_152.default_image_size = resnet_v1.default_image_size
+
+
+def resnet_v1_200(inputs,
+                  num_classes=None,
+                  is_training=True,
+                  global_pool=True,
+                  output_stride=None,
+                  spatial_squeeze=True,
+                  reuse=None,
+                  scope='resnet_v1_200'):
+    """ResNet-200 model of [2]. See resnet_v1() for arg and return description."""
+    blocks = [
+        resnet_utils.Block('block1', bottleneck,
+                           [(256, 64, 1)] * 2 + [(256, 64, 2)]),
+        resnet_utils.Block('block2', bottleneck,
+                           [(512, 128, 1)] * 23 + [(512, 128, 2)]),
+        resnet_utils.Block('block3', bottleneck,
+                           [(1024, 256, 1)] * 35 + [(1024, 256, 2)]),
+        resnet_utils.Block('block4', bottleneck, [(2048, 512, 1)] * 3)
+    ]
+    return resnet_v1(
+        inputs,
+        blocks,
+        num_classes,
+        is_training,
+        global_pool=global_pool,
+        output_stride=output_stride,
+        include_root_block=True,
+        spatial_squeeze=spatial_squeeze,
+        reuse=reuse,
+        scope=scope)
+
+
+resnet_v1_200.default_image_size = resnet_v1.default_image_size
+
+if __name__ == '__main__':
+    input = tf.placeholder(tf.float32, shape=(None, 224, 224, 3), name='input')
+    with slim.arg_scope(resnet_arg_scope()) as sc:
+        logits = resnet_v1_50(input)

modelscope/pipelines/cv/ocr_utils/resnet_utils.py

@@ -0,0 +1,231 @@
+"""Contains building blocks for various versions of Residual Networks.
+Residual networks (ResNets) were proposed in:
+  Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
+  Deep Residual Learning for Image Recognition. arXiv:1512.03385, 2015
+More variants were introduced in:
+  Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
+  Identity Mappings in Deep Residual Networks. arXiv: 1603.05027, 2016
+We can obtain different ResNet variants by changing the network depth, width,
+and form of residual unit. This module implements the infrastructure for
+building them. Concrete ResNet units and full ResNet networks are implemented in
+the accompanying resnet_v1.py and resnet_v2.py modules.
+Compared to https://github.com/KaimingHe/deep-residual-networks, in the current
+implementation we subsample the output activations in the last residual unit of
+each block, instead of subsampling the input activations in the first residual
+unit of each block. The two implementations give identical results but our
+implementation is more memory efficient.
+"""
+
+import collections
+
+import tensorflow as tf
+import tf_slim as slim
+
+if tf.__version__ >= '2.0':
+    tf = tf.compat.v1
+
+
+class Block(collections.namedtuple('Block', ['scope', 'unit_fn', 'args'])):
+    """A named tuple describing a ResNet block.
+    Its parts are:
+      scope: The scope of the `Block`.
+      unit_fn: The ResNet unit function which takes as input a `Tensor` and
+        returns another `Tensor` with the output of the ResNet unit.
+      args: A list of length equal to the number of units in the `Block`. The list
+        contains one (depth, depth_bottleneck, stride) tuple for each unit in the
+        block to serve as argument to unit_fn.
+    """
+
+
+def subsample(inputs, factor, scope=None):
+    """Subsamples the input along the spatial dimensions.
+    Args:
+      inputs: A `Tensor` of size [batch, height_in, width_in, channels].
+      factor: The subsampling factor.
+      scope: Optional variable_scope.
+    Returns:
+      output: A `Tensor` of size [batch, height_out, width_out, channels] with the
+        input, either intact (if factor == 1) or subsampled (if factor > 1).
+    """
+    if factor == 1:
+        return inputs
+    else:
+        return slim.max_pool2d(inputs, [1, 1], stride=factor, scope=scope)
+
+
+def conv2d_same(inputs, num_outputs, kernel_size, stride, rate=1, scope=None):
+    """Strided 2-D convolution with 'SAME' padding.
+    When stride > 1, then we do explicit zero-padding, followed by conv2d with
+    'VALID' padding.
+    Note that
+       net = conv2d_same(inputs, num_outputs, 3, stride=stride)
+    is equivalent to
+       net = slim.conv2d(inputs, num_outputs, 3, stride=1, padding='SAME')
+       net = subsample(net, factor=stride)
+    whereas
+       net = slim.conv2d(inputs, num_outputs, 3, stride=stride, padding='SAME')
+    is different when the input's height or width is even, which is why we add the
+    current function. For more details, see ResnetUtilsTest.testConv2DSameEven().
+    Args:
+      inputs: A 4-D tensor of size [batch, height_in, width_in, channels].
+      num_outputs: An integer, the number of output filters.
+      kernel_size: An int with the kernel_size of the filters.
+      stride: An integer, the output stride.
+      rate: An integer, rate for atrous convolution.
+      scope: Scope.
+    Returns:
+      output: A 4-D tensor of size [batch, height_out, width_out, channels] with
+        the convolution output.
+    """
+    if stride == 1:
+        return slim.conv2d(
+            inputs,
+            num_outputs,
+            kernel_size,
+            stride=1,
+            rate=rate,
+            padding='SAME',
+            scope=scope)
+    else:
+        kernel_size_effective = kernel_size + (kernel_size - 1) * (rate - 1)
+        pad_total = kernel_size_effective - 1
+        pad_beg = pad_total // 2
+        pad_end = pad_total - pad_beg
+        inputs = tf.pad(
+            inputs, [[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]])
+        return slim.conv2d(
+            inputs,
+            num_outputs,
+            kernel_size,
+            stride=stride,
+            rate=rate,
+            padding='VALID',
+            scope=scope)
+
+
+@slim.add_arg_scope
+def stack_blocks_dense(net,
+                       blocks,
+                       output_stride=None,
+                       outputs_collections=None):
+    """Stacks ResNet `Blocks` and controls output feature density.
+    First, this function creates scopes for the ResNet in the form of
+    'block_name/unit_1', 'block_name/unit_2', etc.
+    Second, this function allows the user to explicitly control the ResNet
+    output_stride, which is the ratio of the input to output spatial resolution.
+    This is useful for dense prediction tasks such as semantic segmentation or
+    object detection.
+    Most ResNets consist of 4 ResNet blocks and subsample the activations by a
+    factor of 2 when transitioning between consecutive ResNet blocks. This results
+    to a nominal ResNet output_stride equal to 8. If we set the output_stride to
+    half the nominal network stride (e.g., output_stride=4), then we compute
+    responses twice.
+    Control of the output feature density is implemented by atrous convolution.
+    Args:
+      net: A `Tensor` of size [batch, height, width, channels].
+      blocks: A list of length equal to the number of ResNet `Blocks`. Each
+        element is a ResNet `Block` object describing the units in the `Block`.
+      output_stride: If `None`, then the output will be computed at the nominal
+        network stride. If output_stride is not `None`, it specifies the requested
+        ratio of input to output spatial resolution, which needs to be equal to
+        the product of unit strides from the start up to some level of the ResNet.
+        For example, if the ResNet employs units with strides 1, 2, 1, 3, 4, 1,
+        then valid values for the output_stride are 1, 2, 6, 24 or None (which
+        is equivalent to output_stride=24).
+      outputs_collections: Collection to add the ResNet block outputs.
+    Returns:
+      net: Output tensor with stride equal to the specified output_stride.
+    Raises:
+      ValueError: If the target output_stride is not valid.
+    """
+    # The current_stride variable keeps track of the effective stride of the
+    # activations. This allows us to invoke atrous convolution whenever applying
+    # the next residual unit would result in the activations having stride larger
+    # than the target output_stride.
+    current_stride = 1
+
+    # The atrous convolution rate parameter.
+    rate = 1
+
+    for block in blocks:
+        with tf.variable_scope(block.scope, 'block', [net]):
+            for i, unit in enumerate(block.args):
+                if output_stride is not None and current_stride > output_stride:
+                    raise ValueError(
+                        'The target output_stride cannot be reached.')
+
+                with tf.variable_scope(
+                        'unit_%d' % (i + 1), values=[net]) as sc:
+                    unit_depth, unit_depth_bottleneck, unit_stride = unit
+                    # If we have reached the target output_stride, then we need to employ
+                    # atrous convolution with stride=1 and multiply the atrous rate by the
+                    # current unit's stride for use in subsequent layers.
+                    if output_stride is not None and current_stride == output_stride:
+                        net = block.unit_fn(
+                            net,
+                            depth=unit_depth,
+                            depth_bottleneck=unit_depth_bottleneck,
+                            stride=1,
+                            rate=rate)
+                        rate *= unit_stride
+
+                    else:
+                        net = block.unit_fn(
+                            net,
+                            depth=unit_depth,
+                            depth_bottleneck=unit_depth_bottleneck,
+                            stride=unit_stride,
+                            rate=1)
+                        current_stride *= unit_stride
+                    net = slim.utils.collect_named_outputs(
+                        outputs_collections, sc.name, net)
+
+    if output_stride is not None and current_stride != output_stride:
+        raise ValueError('The target output_stride cannot be reached.')
+
+    return net
+
+
+def resnet_arg_scope(weight_decay=0.0001,
+                     batch_norm_decay=0.997,
+                     batch_norm_epsilon=1e-5,
+                     batch_norm_scale=True):
+    """Defines the default ResNet arg scope.
+    TODO(gpapan): The batch-normalization related default values above are
+      appropriate for use in conjunction with the reference ResNet models
+      released at https://github.com/KaimingHe/deep-residual-networks. When
+      training ResNets from scratch, they might need to be tuned.
+    Args:
+      weight_decay: The weight decay to use for regularizing the model.
+      batch_norm_decay: The moving average decay when estimating layer activation
+        statistics in batch normalization.
+      batch_norm_epsilon: Small constant to prevent division by zero when
+        normalizing activations by their variance in batch normalization.
+      batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
+        activations in the batch normalization layer.
+    Returns:
+      An `arg_scope` to use for the resnet models.
+    """
+    batch_norm_params = {
+        'decay': batch_norm_decay,
+        'epsilon': batch_norm_epsilon,
+        'scale': batch_norm_scale,
+        'updates_collections': tf.GraphKeys.UPDATE_OPS,
+    }
+
+    with slim.arg_scope(
+        [slim.conv2d],
+            weights_regularizer=slim.l2_regularizer(weight_decay),
+            weights_initializer=slim.variance_scaling_initializer(),
+            activation_fn=tf.nn.relu,
+            normalizer_fn=slim.batch_norm,
+            normalizer_params=batch_norm_params):
+        with slim.arg_scope([slim.batch_norm], **batch_norm_params):
+            # The following implies padding='SAME' for pool1, which makes feature
+            # alignment easier for dense prediction tasks. This is also used in
+            # https://github.com/facebook/fb.resnet.torch. However the accompanying
+            # code of 'Deep Residual Learning for Image Recognition' uses
+            # padding='VALID' for pool1. You can switch to that choice by setting
+            # slim.arg_scope([slim.max_pool2d], padding='VALID').
+            with slim.arg_scope([slim.max_pool2d], padding='VALID') as arg_sc:
+                return arg_sc

modelscope/pipelines/cv/ocr_utils/utils.py

@@ -0,0 +1,108 @@
+import cv2
+import numpy as np
+
+
+def rboxes_to_polygons(rboxes):
+    """
+    Convert rboxes to polygons
+    ARGS
+        `rboxes`: [n, 5]
+    RETURN
+        `polygons`: [n, 8]
+    """
+
+    theta = rboxes[:, 4:5]
+    cxcy = rboxes[:, :2]
+    half_w = rboxes[:, 2:3] / 2.
+    half_h = rboxes[:, 3:4] / 2.
+    v1 = np.hstack([np.cos(theta) * half_w, np.sin(theta) * half_w])
+    v2 = np.hstack([-np.sin(theta) * half_h, np.cos(theta) * half_h])
+    p1 = cxcy - v1 - v2
+    p2 = cxcy + v1 - v2
+    p3 = cxcy + v1 + v2
+    p4 = cxcy - v1 + v2
+    polygons = np.hstack([p1, p2, p3, p4])
+    return polygons
+
+
+def cal_width(box):
+    pd1 = point_dist(box[0], box[1], box[2], box[3])
+    pd2 = point_dist(box[4], box[5], box[6], box[7])
+    return (pd1 + pd2) / 2
+
+
+def point_dist(x1, y1, x2, y2):
+    return np.sqrt((x2 - x1) * (x2 - x1) + (y2 - y1) * (y2 - y1))
+
+
+def draw_polygons(img, polygons):
+    for p in polygons.tolist():
+        p = [int(o) for o in p]
+        cv2.line(img, (p[0], p[1]), (p[2], p[3]), (0, 255, 0), 1)
+        cv2.line(img, (p[2], p[3]), (p[4], p[5]), (0, 255, 0), 1)
+        cv2.line(img, (p[4], p[5]), (p[6], p[7]), (0, 255, 0), 1)
+        cv2.line(img, (p[6], p[7]), (p[0], p[1]), (0, 255, 0), 1)
+    return img
+
+
+def nms_python(boxes):
+    boxes = sorted(boxes, key=lambda x: -x[8])
+    nms_flag = [True] * len(boxes)
+    for i, a in enumerate(boxes):
+        if not nms_flag[i]:
+            continue
+        else:
+            for j, b in enumerate(boxes):
+                if not j > i:
+                    continue
+                if not nms_flag[j]:
+                    continue
+                score_a = a[8]
+                score_b = b[8]
+                rbox_a = polygon2rbox(a[:8])
+                rbox_b = polygon2rbox(b[:8])
+                if point_in_rbox(rbox_a[:2], rbox_b) or point_in_rbox(
+                        rbox_b[:2], rbox_a):
+                    if score_a > score_b:
+                        nms_flag[j] = False
+    boxes_nms = []
+    for i, box in enumerate(boxes):
+        if nms_flag[i]:
+            boxes_nms.append(box)
+    return boxes_nms
+
+
+def point_in_rbox(c, rbox):
+    cx0, cy0 = c[0], c[1]
+    cx1, cy1 = rbox[0], rbox[1]
+    w, h = rbox[2], rbox[3]
+    theta = rbox[4]
+    dist_x = np.abs((cx1 - cx0) * np.cos(theta) + (cy1 - cy0) * np.sin(theta))
+    dist_y = np.abs(-(cx1 - cx0) * np.sin(theta) + (cy1 - cy0) * np.cos(theta))
+    return ((dist_x < w / 2.0) and (dist_y < h / 2.0))
+
+
+def polygon2rbox(polygon):
+    x1, x2, x3, x4 = polygon[0], polygon[2], polygon[4], polygon[6]
+    y1, y2, y3, y4 = polygon[1], polygon[3], polygon[5], polygon[7]
+    c_x = (x1 + x2 + x3 + x4) / 4
+    c_y = (y1 + y2 + y3 + y4) / 4
+    w1 = point_dist(x1, y1, x2, y2)
+    w2 = point_dist(x3, y3, x4, y4)
+    h1 = point_line_dist(c_x, c_y, x1, y1, x2, y2)
+    h2 = point_line_dist(c_x, c_y, x3, y3, x4, y4)
+    h = h1 + h2
+    w = (w1 + w2) / 2
+    theta1 = np.arctan2(y2 - y1, x2 - x1)
+    theta2 = np.arctan2(y3 - y4, x3 - x4)
+    theta = (theta1 + theta2) / 2.0
+    return [c_x, c_y, w, h, theta]
+
+
+def point_line_dist(px, py, x1, y1, x2, y2):
+    eps = 1e-6
+    dx = x2 - x1
+    dy = y2 - y1
+    div = np.sqrt(dx * dx + dy * dy) + eps
+    dist = np.abs(px * dy - py * dx + x2 * y1 - y2 * x1) / div
+    return dist

modelscope/pipelines/multi_modal/__init__.py

@@ -1 +1 @@
-from .image_captioning import ImageCaptionPipeline
+from .image_captioning_pipeline import ImageCaptionPipeline

modelscope/pipelines/multi_modal/image_captioning_pipeline.py

@@ -0,0 +1,33 @@
+from typing import Any, Dict, Union
+
+from modelscope.preprocessors import OfaImageCaptionPreprocessor, Preprocessor
+from modelscope.utils.constant import Tasks
+from modelscope.utils.logger import get_logger
+from ..base import Model, Pipeline
+from ..builder import PIPELINES
+
+logger = get_logger()
+
+
+@PIPELINES.register_module(Tasks.image_captioning, module_name='ofa')
+class ImageCaptionPipeline(Pipeline):
+
+    def __init__(self,
+                 model: Union[Model, str],
+                 preprocessor: [Preprocessor] = None,
+                 **kwargs):
+        super().__init__()
+        assert isinstance(model, str) or isinstance(model, Model), \
+            'model must be a single str or OfaForImageCaptioning'
+        if isinstance(model, str):
+            pipe_model = Model.from_pretrained(model)
+        elif isinstance(model, Model):
+            pipe_model = model
+        else:
+            raise NotImplementedError
+        if preprocessor is None and pipe_model:
+            preprocessor = OfaImageCaptionPreprocessor(model_dir=model)
+        super().__init__(model=pipe_model, preprocessor=preprocessor, **kwargs)
+
+    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        return inputs

modelscope/pipelines/nlp/__init__.py

@@ -1,4 +1,5 @@
 from .sentence_similarity_pipeline import *  # noqa F403
 from .sequence_classification_pipeline import *  # noqa F403
 from .text_generation_pipeline import *  # noqa F403
+from .word_segmentation_pipeline import *  # noqa F403
 from .zero_shot_classification_pipeline import *  # noqa F403

modelscope/pipelines/nlp/sentence_similarity_pipeline.py

@@ -1,8 +1,5 @@
-import os
-import uuid
 from typing import Any, Dict, Union
 
-import json
 import numpy as np
 
 from modelscope.models.nlp import SbertForSentenceSimilarity

modelscope/pipelines/nlp/sequence_classification_pipeline.py

@@ -1,8 +1,5 @@
-import os
-import uuid
 from typing import Any, Dict, Union
 
-import json
 import numpy as np
 
 from modelscope.models.nlp import BertForSequenceClassification

modelscope/pipelines/nlp/text_generation_pipeline.py

@@ -1,7 +1,7 @@
 from typing import Dict, Optional, Union
 
 from modelscope.models import Model
-from modelscope.models.nlp import PalmForTextGenerationModel
+from modelscope.models.nlp import PalmForTextGeneration
 from modelscope.preprocessors import TextGenerationPreprocessor
 from modelscope.utils.constant import Tasks
 from ..base import Pipeline, Tensor
@@ -10,11 +10,11 @@ from ..builder import PIPELINES
 __all__ = ['TextGenerationPipeline']
 
 
-@PIPELINES.register_module(Tasks.text_generation, module_name=r'palm')
+@PIPELINES.register_module(Tasks.text_generation, module_name=r'palm2.0')
 class TextGenerationPipeline(Pipeline):
 
     def __init__(self,
-                 model: Union[PalmForTextGenerationModel, str],
+                 model: Union[PalmForTextGeneration, str],
                  preprocessor: Optional[TextGenerationPreprocessor] = None,
                  **kwargs):
         """use `model` and `preprocessor` to create a nlp text classification pipeline for prediction
@@ -23,16 +23,16 @@ class TextGenerationPipeline(Pipeline):
             model (SequenceClassificationModel): a model instance
             preprocessor (SequenceClassificationPreprocessor): a preprocessor instance
         """
-        sc_model = model if isinstance(
-            model,
-            PalmForTextGenerationModel) else Model.from_pretrained(model)
+        model = model if isinstance(
+            model, PalmForTextGeneration) else Model.from_pretrained(model)
         if preprocessor is None:
             preprocessor = TextGenerationPreprocessor(
-                sc_model.model_dir,
+                model.model_dir,
+                model.tokenizer,
                 first_sequence='sentence',
                 second_sequence=None)
-        super().__init__(model=sc_model, preprocessor=preprocessor, **kwargs)
-        self.tokenizer = preprocessor.tokenizer
+        super().__init__(model=model, preprocessor=preprocessor, **kwargs)
+        self.tokenizer = model.tokenizer
 
     def postprocess(self, inputs: Dict[str, Tensor]) -> Dict[str, str]:
         """process the prediction results
@@ -43,17 +43,20 @@ class TextGenerationPipeline(Pipeline):
         Returns:
             Dict[str, str]: the prediction results
         """
+        replace_tokens_bert = (('[unused0]', ''), ('[PAD]', ''),
+                               ('[unused1]', ''), (r' +', ' '), ('[SEP]', ''),
+                               ('[unused2]', ''), ('[CLS]', ''), ('[UNK]', ''))
+        replace_tokens_roberta = ((r' +', ' '), ('<mask>', '<q>'), ('<pad>',
+                                                                    ''),
+                                  ('<s>', ''), ('</s>', ''), ('<unk>', ' '))
 
-        vocab_size = len(self.tokenizer.vocab)
         pred_list = inputs['predictions']
         pred_ids = pred_list[0][0].cpu().numpy().tolist()
-        for j in range(len(pred_ids)):
-            if pred_ids[j] >= vocab_size:
-                pred_ids[j] = 100
-        pred = self.tokenizer.convert_ids_to_tokens(pred_ids)
-        pred_string = ''.join(pred).replace(
-            '##',
-            '').split('[SEP]')[0].replace('[CLS]',
-                                          '').replace('[SEP]',
-                                                      '').replace('[UNK]', '')
+        pred_string = self.tokenizer.decode(pred_ids)
+        for _old, _new in replace_tokens_bert:
+            pred_string = pred_string.replace(_old, _new)
+        pred_string.strip()
+        for _old, _new in replace_tokens_roberta:
+            pred_string = pred_string.replace(_old, _new)
+        pred_string.strip()
         return {'text': pred_string}

modelscope/pipelines/nlp/word_segmentation_pipeline.py

@@ -0,0 +1,69 @@
+from typing import Any, Dict, Optional, Union
+
+from modelscope.models import Model
+from modelscope.models.nlp import StructBertForTokenClassification
+from modelscope.preprocessors import TokenClassifcationPreprocessor
+from modelscope.utils.constant import Tasks
+from ..base import Pipeline, Tensor
+from ..builder import PIPELINES
+
+__all__ = ['WordSegmentationPipeline']
+
+
+@PIPELINES.register_module(
+    Tasks.word_segmentation,
+    module_name=r'structbert-chinese-word-segmentation')
+class WordSegmentationPipeline(Pipeline):
+
+    def __init__(self,
+                 model: Union[StructBertForTokenClassification, str],
+                 preprocessor: Optional[TokenClassifcationPreprocessor] = None,
+                 **kwargs):
+        """use `model` and `preprocessor` to create a nlp word segmentation pipeline for prediction
+
+        Args:
+            model (StructBertForTokenClassification): a model instance
+            preprocessor (TokenClassifcationPreprocessor): a preprocessor instance
+        """
+        model = model if isinstance(
+            model,
+            StructBertForTokenClassification) else Model.from_pretrained(model)
+        if preprocessor is None:
+            preprocessor = TokenClassifcationPreprocessor(model.model_dir)
+        super().__init__(model=model, preprocessor=preprocessor, **kwargs)
+        self.tokenizer = preprocessor.tokenizer
+        self.config = model.config
+        self.id2label = self.config.id2label
+
+    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, str]:
+        """process the prediction results
+
+        Args:
+            inputs (Dict[str, Any]): _description_
+
+        Returns:
+            Dict[str, str]: the prediction results
+        """
+
+        pred_list = inputs['predictions']
+        labels = []
+        for pre in pred_list:
+            labels.append(self.id2label[pre])
+        labels = labels[1:-1]
+        chunks = []
+        chunk = ''
+        assert len(inputs['text']) == len(labels)
+        for token, label in zip(inputs['text'], labels):
+            if label[0] == 'B' or label[0] == 'I':
+                chunk += token
+            else:
+                chunk += token
+                chunks.append(chunk)
+                chunk = ''
+        if chunk:
+            chunks.append(chunk)
+        seg_result = ' '.join(chunks)
+        rst = {
+            'output': seg_result,
+        }
+        return rst

modelscope/pipelines/outputs.py

@@ -54,6 +54,13 @@ TASK_OUTPUTS = {
     # }
     Tasks.pose_estimation: ['poses', 'boxes'],
 
+    # ocr detection result for single sample
+    # {
+    #   "det_polygons": np.array with shape [num_text, 8], each box is
+    #       [x1, y1, x2, y2, x3, y3, x4, y4]
+    # }
+    Tasks.ocr_detection: ['det_polygons'],
+
     # ============ nlp tasks ===================
 
     # text classification result for single sample
@@ -69,8 +76,27 @@ TASK_OUTPUTS = {
     # }
     Tasks.text_generation: ['text'],
 
+    # word segmentation result for single sample
+    # {
+    #   "output": "今天 天气 不错 ， 适合 出去 游玩"
+    # }
+    Tasks.word_segmentation: ['output'],
+
+    # sentence similarity result for single sample
+    #   {
+    #       "labels": "1",
+    #       "scores": 0.9
+    #   }
+    Tasks.sentence_similarity: ['scores', 'labels'],
+
     # ============ audio tasks ===================
 
+    # audio processed for single file in PCM format
+    # {
+    #   "output_pcm": np.array with shape(samples,) and dtype float32
+    # }
+    Tasks.speech_signal_process: ['output_pcm'],
+
     # ============ multi-modal tasks ===================
 
     # image caption result for single sample

modelscope/preprocessors/__init__.py

@@ -1,7 +1,10 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
+from .audio import LinearAECAndFbank
 from .base import Preprocessor
 from .builder import PREPROCESSORS, build_preprocessor
 from .common import Compose
 from .image import LoadImage, load_image
+from .multi_model import OfaImageCaptionPreprocessor
 from .nlp import *  # noqa F403
+from .text_to_speech import *  # noqa F403

modelscope/preprocessors/audio.py

@@ -0,0 +1,231 @@
+import ctypes
+import os
+from typing import Any, Dict
+
+import numpy as np
+import scipy.io.wavfile as wav
+import torch
+from numpy.ctypeslib import ndpointer
+
+from modelscope.utils.constant import Fields
+from .builder import PREPROCESSORS
+
+
+def load_wav(path):
+    samp_rate, data = wav.read(path)
+    return np.float32(data), samp_rate
+
+
+def load_library(libaec):
+    libaec_in_cwd = os.path.join('.', libaec)
+    if os.path.exists(libaec_in_cwd):
+        libaec = libaec_in_cwd
+    mitaec = ctypes.cdll.LoadLibrary(libaec)
+    fe_process = mitaec.fe_process_inst
+    fe_process.argtypes = [
+        ndpointer(ctypes.c_float, flags='C_CONTIGUOUS'),
+        ndpointer(ctypes.c_float, flags='C_CONTIGUOUS'), ctypes.c_int,
+        ndpointer(ctypes.c_float, flags='C_CONTIGUOUS'),
+        ndpointer(ctypes.c_float, flags='C_CONTIGUOUS'),
+        ndpointer(ctypes.c_float, flags='C_CONTIGUOUS')
+    ]
+    return fe_process
+
+
+def do_linear_aec(fe_process, mic, ref, int16range=True):
+    mic = np.float32(mic)
+    ref = np.float32(ref)
+    if len(mic) > len(ref):
+        mic = mic[:len(ref)]
+    out_mic = np.zeros_like(mic)
+    out_linear = np.zeros_like(mic)
+    out_echo = np.zeros_like(mic)
+    out_ref = np.zeros_like(mic)
+    if int16range:
+        mic /= 32768
+        ref /= 32768
+    fe_process(mic, ref, len(mic), out_mic, out_linear, out_echo)
+    # out_ref not in use here
+    if int16range:
+        out_mic *= 32768
+        out_linear *= 32768
+        out_echo *= 32768
+    return out_mic, out_ref, out_linear, out_echo
+
+
+def load_kaldi_feature_transform(filename):
+    fp = open(filename, 'r')
+    all_str = fp.read()
+    pos1 = all_str.find('AddShift')
+    pos2 = all_str.find('[', pos1)
+    pos3 = all_str.find(']', pos2)
+    mean = np.fromstring(all_str[pos2 + 1:pos3], dtype=np.float32, sep=' ')
+    pos1 = all_str.find('Rescale')
+    pos2 = all_str.find('[', pos1)
+    pos3 = all_str.find(']', pos2)
+    scale = np.fromstring(all_str[pos2 + 1:pos3], dtype=np.float32, sep=' ')
+    fp.close()
+    return mean, scale
+
+
+class Feature:
+    r"""Extract feat from one utterance.
+    """
+
+    def __init__(self,
+                 fbank_config,
+                 feat_type='spec',
+                 mvn_file=None,
+                 cuda=False):
+        r"""
+
+        Args:
+            fbank_config (dict):
+            feat_type (str):
+                raw: do nothing
+                fbank: use kaldi.fbank
+                spec: Real/Imag
+                logpow: log(1+|x|^2)
+            mvn_file (str): the path of data file for mean variance normalization
+            cuda:
+        """
+        self.fbank_config = fbank_config
+        self.feat_type = feat_type
+        self.n_fft = fbank_config['frame_length'] * fbank_config[
+            'sample_frequency'] // 1000
+        self.hop_length = fbank_config['frame_shift'] * fbank_config[
+            'sample_frequency'] // 1000
+        self.window = torch.hamming_window(self.n_fft, periodic=False)
+
+        self.mvn = False
+        if mvn_file is not None and os.path.exists(mvn_file):
+            print(f'loading mvn file: {mvn_file}')
+            shift, scale = load_kaldi_feature_transform(mvn_file)
+            self.shift = torch.from_numpy(shift)
+            self.scale = torch.from_numpy(scale)
+            self.mvn = True
+        if cuda:
+            self.window = self.window.cuda()
+            if self.mvn:
+                self.shift = self.shift.cuda()
+                self.scale = self.scale.cuda()
+
+    def compute(self, utt):
+        r"""
+
+        Args:
+            utt: in [-32768, 32767] range
+
+        Returns:
+             [..., T, F]
+        """
+        if self.feat_type == 'raw':
+            return utt
+        elif self.feat_type == 'fbank':
+            # have to use local import before modelscope framework supoort lazy loading
+            import torchaudio.compliance.kaldi as kaldi
+            if len(utt.shape) == 1:
+                utt = utt.unsqueeze(0)
+            feat = kaldi.fbank(utt, **self.fbank_config)
+        elif self.feat_type == 'spec':
+            spec = torch.stft(
+                utt / 32768,
+                self.n_fft,
+                self.hop_length,
+                self.n_fft,
+                self.window,
+                center=False,
+                return_complex=True)
+            feat = torch.cat([spec.real, spec.imag], dim=-2).permute(-1, -2)
+        elif self.feat_type == 'logpow':
+            spec = torch.stft(
+                utt,
+                self.n_fft,
+                self.hop_length,
+                self.n_fft,
+                self.window,
+                center=False,
+                return_complex=True)
+            abspow = torch.abs(spec)**2
+            feat = torch.log(1 + abspow).permute(-1, -2)
+        return feat
+
+    def normalize(self, feat):
+        if self.mvn:
+            feat = feat + self.shift
+            feat = feat * self.scale
+        return feat
+
+
+@PREPROCESSORS.register_module(Fields.audio)
+class LinearAECAndFbank:
+    SAMPLE_RATE = 16000
+
+    def __init__(self, io_config):
+        self.trunc_length = 7200 * self.SAMPLE_RATE
+        self.linear_aec_delay = io_config['linear_aec_delay']
+        self.feature = Feature(io_config['fbank_config'],
+                               io_config['feat_type'], io_config['mvn'])
+        self.mitaec = load_library(io_config['mitaec_library'])
+        self.mask_on_mic = io_config['mask_on'] == 'nearend_mic'
+
+    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
+        """ linear filtering the near end mic and far end audio, then extract the feature
+        :param data: dict with two keys and correspond audios: "nearend_mic" and "farend_speech"
+        :return: dict with two keys and Tensor values: "base" linear filtered audio，and "feature"
+        """
+        # read files
+        nearend_mic, fs = load_wav(data['nearend_mic'])
+        assert fs == self.SAMPLE_RATE, f'The sample rate should be {self.SAMPLE_RATE}'
+        farend_speech, fs = load_wav(data['farend_speech'])
+        assert fs == self.SAMPLE_RATE, f'The sample rate should be {self.SAMPLE_RATE}'
+        if 'nearend_speech' in data:
+            nearend_speech, fs = load_wav(data['nearend_speech'])
+            assert fs == self.SAMPLE_RATE, f'The sample rate should be {self.SAMPLE_RATE}'
+        else:
+            nearend_speech = np.zeros_like(nearend_mic)
+
+        out_mic, out_ref, out_linear, out_echo = do_linear_aec(
+            self.mitaec, nearend_mic, farend_speech)
+        # fix 20ms linear aec delay by delaying the target speech
+        extra_zeros = np.zeros([int(self.linear_aec_delay * fs)])
+        nearend_speech = np.concatenate([extra_zeros, nearend_speech])
+        # truncate files to the same length
+        flen = min(
+            len(out_mic), len(out_ref), len(out_linear), len(out_echo),
+            len(nearend_speech))
+        fstart = 0
+        flen = min(flen, self.trunc_length)
+        nearend_mic, out_ref, out_linear, out_echo, nearend_speech = (
+            out_mic[fstart:flen], out_ref[fstart:flen],
+            out_linear[fstart:flen], out_echo[fstart:flen],
+            nearend_speech[fstart:flen])
+
+        # extract features (frames, [mic, linear, ref, aes?])
+        feat = torch.FloatTensor()
+
+        nearend_mic = torch.from_numpy(np.float32(nearend_mic))
+        fbank_nearend_mic = self.feature.compute(nearend_mic)
+        feat = torch.cat([feat, fbank_nearend_mic], dim=1)
+
+        out_linear = torch.from_numpy(np.float32(out_linear))
+        fbank_out_linear = self.feature.compute(out_linear)
+        feat = torch.cat([feat, fbank_out_linear], dim=1)
+
+        out_echo = torch.from_numpy(np.float32(out_echo))
+        fbank_out_echo = self.feature.compute(out_echo)
+        feat = torch.cat([feat, fbank_out_echo], dim=1)
+
+        # feature transform
+        feat = self.feature.normalize(feat)
+
+        # prepare target
+        if nearend_speech is not None:
+            nearend_speech = torch.from_numpy(np.float32(nearend_speech))
+
+        if self.mask_on_mic:
+            base = nearend_mic
+        else:
+            base = out_linear
+        out_data = {'base': base, 'target': nearend_speech, 'feature': feat}
+        return out_data

modelscope/preprocessors/multi_model.py

@@ -1,32 +1,50 @@
-from typing import Any, Dict
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import os.path as osp
+from typing import Any, Dict, Union
 
 import numpy as np
 import torch
+from maas_hub.snapshot_download import snapshot_download
 from PIL import Image
 
-from modelscope.pipelines.base import Input
-from modelscope.preprocessors import load_image
-from modelscope.utils.constant import Tasks
-from modelscope.utils.logger import get_logger
-from ..base import Pipeline
-from ..builder import PIPELINES
+from modelscope.utils.constant import Fields, ModelFile
+from modelscope.utils.hub import get_model_cache_dir
+from modelscope.utils.type_assert import type_assert
+from .base import Preprocessor
+from .builder import PREPROCESSORS
+from .image import load_image
 
-logger = get_logger()
+__all__ = [
+    'OfaImageCaptionPreprocessor',
+]
 
 
-@PIPELINES.register_module(Tasks.image_captioning, module_name='ofa')
-class ImageCaptionPipeline(Pipeline):
-    # TODO: refine using modelhub
-    def __init__(self, model: str, bpe_dir: str):
-        super().__init__()
-        # turn on cuda if GPU is available
+@PREPROCESSORS.register_module(
+    Fields.multi_modal, module_name=r'ofa-image-caption')
+class OfaImageCaptionPreprocessor(Preprocessor):
+
+    def __init__(self, model_dir: str, *args, **kwargs):
+        """preprocess the data via the vocab.txt from the `model_dir` path
+
+        Args:
+            model_dir (str): model path
+        """
+        super().__init__(*args, **kwargs)
+
+        if osp.exists(model_dir):
+            local_model_dir = model_dir
+        else:
+            cache_path = get_model_cache_dir(model_dir)
+            local_model_dir = cache_path if osp.exists(
+                cache_path) else snapshot_download(model_dir)
+        local_model = osp.join(local_model_dir, ModelFile.TORCH_MODEL_FILE)
+        bpe_dir = local_model_dir
+
         from fairseq import checkpoint_utils, tasks, utils
         from ofa.tasks.mm_tasks import CaptionTask
 
         tasks.register_task('caption', CaptionTask)
-        use_cuda = False
-        # use fp16 only when GPU is available
-        use_fp16 = False
+
         overrides = {
             'bpe_dir': bpe_dir,
             'eval_cider': False,
@@ -35,21 +53,9 @@ class ImageCaptionPipeline(Pipeline):
             'no_repeat_ngram_size': 3,
             'seed': 7
         }
-        models, cfg, task = checkpoint_utils.load_model_ensemble_and_task(
-            utils.split_paths(model), arg_overrides=overrides)
-
-        # Move models to GPU
-        for model in models:
-            model.eval()
-            if use_cuda:
-                model.cuda()
-            if use_fp16:
-                model.half()
-            model.prepare_for_inference_(cfg)
-        self.models = models
-        # Initialize generator
-        self.generator = task.build_generator(models, cfg.generation)
-
+        model, cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+            utils.split_paths(local_model), arg_overrides=overrides)
+        del model
         # Initialize transform
         from torchvision import transforms
         mean = [0.5, 0.5, 0.5]
@@ -69,7 +75,8 @@ class ImageCaptionPipeline(Pipeline):
         self.eos_item = torch.LongTensor([task.src_dict.eos()])
         self.pad_idx = task.src_dict.pad()
 
-    def preprocess(self, input: Input) -> Dict[str, Any]:
+    @type_assert(object, (str, tuple))
+    def __call__(self, data: Union[str, tuple]) -> Dict[str, Any]:
 
         def encode_text(text, length=None, append_bos=False, append_eos=False):
             s = self.task.tgt_dict.encode_line(
@@ -88,7 +95,7 @@ class ImageCaptionPipeline(Pipeline):
             patch_image = self.patch_resize_transform(input).unsqueeze(0)
         else:
             patch_image = self.patch_resize_transform(
-                load_image(input)).unsqueeze(0)
+                load_image(data)).unsqueeze(0)
         patch_mask = torch.tensor([True])
         text = 'what does the image describe?'
         src_text = encode_text(
@@ -105,17 +112,3 @@ class ImageCaptionPipeline(Pipeline):
             }
         }
         return sample
-
-    def forward(self, input: Dict[str, Any]) -> Dict[str, Any]:
-        from ofa.utils.eval_utils import eval_caption
-
-        results, _ = eval_caption(self.task, self.generator, self.models,
-                                  input)
-        return {
-            'image_id': results[0]['image_id'],
-            'caption': results[0]['caption']
-        }
-
-    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
-        # What should we do here ?
-        return inputs

modelscope/preprocessors/nlp.py

@@ -12,7 +12,8 @@ from .builder import PREPROCESSORS
 
 __all__ = [
     'Tokenize', 'SequenceClassificationPreprocessor',
-    'TextGenerationPreprocessor', 'ZeroShotClassificationPreprocessor'
+    'TextGenerationPreprocessor', 'ZeroShotClassificationPreprocessor',
+    'TokenClassifcationPreprocessor'
 ]
 
 
@@ -53,12 +54,12 @@ class SequenceClassificationPreprocessor(Preprocessor):
         self.tokenizer = AutoTokenizer.from_pretrained(self.model_dir)
         print(f'this is the tokenzier {self.tokenizer}')
 
-    @type_assert(object, (str, tuple))
-    def __call__(self, data: Union[str, tuple]) -> Dict[str, Any]:
+    @type_assert(object, (str, tuple, Dict))
+    def __call__(self, data: Union[str, tuple, Dict]) -> Dict[str, Any]:
         """process the raw input data
 
         Args:
-            data (str or tuple):
+            data (str or tuple, Dict):
             sentence1 (str): a sentence
                     Example:
                         'you are so handsome.'
@@ -70,22 +71,31 @@ class SequenceClassificationPreprocessor(Preprocessor):
                 sentence2 (str): a sentence
                     Example:
                         'you are so beautiful.'
+            or
+            {field1: field_value1, field2: field_value2}
+            field1 (str): field name, default 'first_sequence'
+            field_value1 (str): a sentence
+                    Example:
+                        'you are so handsome.'
+
+            field2 (str): field name, default 'second_sequence'
+            field_value2 (str): a sentence
+                Example:
+                    'you are so beautiful.'
 
         Returns:
             Dict[str, Any]: the preprocessed data
         """
-
-        if not isinstance(data, tuple):
-            data = (
-                data,
-                None,
-            )
-
-        sentence1, sentence2 = data
-        new_data = {
-            self.first_sequence: sentence1,
-            self.second_sequence: sentence2
-        }
+        if isinstance(data, str):
+            new_data = {self.first_sequence: data}
+        elif isinstance(data, tuple):
+            sentence1, sentence2 = data
+            new_data = {
+                self.first_sequence: sentence1,
+                self.second_sequence: sentence2
+            }
+        else:
+            new_data = data
 
         # preprocess the data for the model input
 
@@ -115,17 +125,15 @@ class SequenceClassificationPreprocessor(Preprocessor):
         return rst
 
 
-@PREPROCESSORS.register_module(Fields.nlp, module_name=r'palm')
+@PREPROCESSORS.register_module(Fields.nlp, module_name=r'palm2.0')
 class TextGenerationPreprocessor(Preprocessor):
 
-    def __init__(self, model_dir: str, *args, **kwargs):
+    def __init__(self, model_dir: str, tokenizer, *args, **kwargs):
         """preprocess the data using the vocab.txt from the `model_dir` path
 
         Args:
             model_dir (str): model path
         """
-        from sofa import PalmTokenizer
-
         super().__init__(*args, **kwargs)
 
         self.model_dir: str = model_dir
@@ -134,7 +142,7 @@ class TextGenerationPreprocessor(Preprocessor):
         self.second_sequence: str = kwargs.pop('second_sequence',
                                                'second_sequence')
         self.sequence_length: int = kwargs.pop('sequence_length', 128)
-        self.tokenizer = PalmTokenizer.from_pretrained(model_dir)
+        self.tokenizer = tokenizer
 
     @type_assert(object, str)
     def __call__(self, data: str) -> Dict[str, Any]:
@@ -153,7 +161,7 @@ class TextGenerationPreprocessor(Preprocessor):
         new_data = {self.first_sequence: data}
         # preprocess the data for the model input
 
-        rst = {'input_ids': [], 'attention_mask': [], 'token_type_ids': []}
+        rst = {'input_ids': [], 'attention_mask': []}
 
         max_seq_length = self.sequence_length
 
@@ -168,7 +176,6 @@ class TextGenerationPreprocessor(Preprocessor):
 
         rst['input_ids'].append(feature['input_ids'])
         rst['attention_mask'].append(feature['attention_mask'])
-        rst['token_type_ids'].append(feature['token_type_ids'])
 
         return {k: torch.tensor(v) for k, v in rst.items()}
 
@@ -191,7 +198,6 @@ class ZeroShotClassificationPreprocessor(Preprocessor):
         self.sequence_length = kwargs.pop('sequence_length', 512)
         self.candidate_labels = kwargs.pop('candidate_labels')
         self.hypothesis_template = kwargs.pop('hypothesis_template', '{}')
-        self.tokenizer = SbertTokenizer.from_pretrained(self.model_dir)
 
     @type_assert(object, str)
     def __call__(self, data: str) -> Dict[str, Any]:
@@ -216,3 +222,52 @@ class ZeroShotClassificationPreprocessor(Preprocessor):
             return_tensors='pt',
             truncation_strategy='only_first')
         return features
+
+
+@PREPROCESSORS.register_module(
+    Fields.nlp, module_name=r'bert-token-classification')
+class TokenClassifcationPreprocessor(Preprocessor):
+
+    def __init__(self, model_dir: str, *args, **kwargs):
+        """preprocess the data via the vocab.txt from the `model_dir` path
+
+        Args:
+            model_dir (str): model path
+        """
+
+        super().__init__(*args, **kwargs)
+
+        from sofa import SbertTokenizer
+        self.model_dir: str = model_dir
+        self.tokenizer = SbertTokenizer.from_pretrained(self.model_dir)
+
+    @type_assert(object, str)
+    def __call__(self, data: str) -> Dict[str, Any]:
+        """process the raw input data
+
+        Args:
+            data (str): a sentence
+                Example:
+                    'you are so handsome.'
+
+        Returns:
+            Dict[str, Any]: the preprocessed data
+        """
+
+        # preprocess the data for the model input
+
+        text = data.replace(' ', '').strip()
+        tokens = []
+        for token in text:
+            token = self.tokenizer.tokenize(token)
+            tokens.extend(token)
+        input_ids = self.tokenizer.convert_tokens_to_ids(tokens)
+        input_ids = self.tokenizer.build_inputs_with_special_tokens(input_ids)
+        attention_mask = [1] * len(input_ids)
+        token_type_ids = [0] * len(input_ids)
+        return {
+            'text': text,
+            'input_ids': input_ids,
+            'attention_mask': attention_mask,
+            'token_type_ids': token_type_ids
+        }

modelscope/preprocessors/text_to_speech.py

@@ -0,0 +1,51 @@
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import io
+from typing import Any, Dict, Union
+
+from modelscope.fileio import File
+from modelscope.models.audio.tts.frontend import GenericTtsFrontend
+from modelscope.models.base import Model
+from modelscope.utils.audio.tts_exceptions import *  # noqa F403
+from modelscope.utils.constant import Fields
+from .base import Preprocessor
+from .builder import PREPROCESSORS
+
+__all__ = ['TextToTacotronSymbols', 'text_to_tacotron_symbols']
+
+
+@PREPROCESSORS.register_module(
+    Fields.audio, module_name=r'text_to_tacotron_symbols')
+class TextToTacotronSymbols(Preprocessor):
+    """extract tacotron symbols from text.
+
+    Args:
+        res_path (str): TTS frontend resource url
+        lang_type (str): language type, valid values are "pinyin" and "chenmix"
+    """
+
+    def __init__(self, model_name, lang_type='pinyin'):
+        self._frontend_model = Model.from_pretrained(
+            model_name, lang_type=lang_type)
+        assert self._frontend_model is not None, 'load model from pretained failed'
+
+    def __call__(self, data: str) -> Dict[str, Any]:
+        """Call functions to load text and get tacotron symbols.
+
+        Args:
+            input (str): text with utf-8
+        Returns:
+            symbos (list[str]): texts in tacotron symbols format.
+        """
+        return self._frontend_model.forward(data)
+
+
+def text_to_tacotron_symbols(text='', path='./', lang='pinyin'):
+    """ simple interface to transform text to tacotron symbols
+
+    Args:
+        text (str): input text
+        path (str): resource path
+        lang (str): language type from one of "pinyin" and "chenmix"
+    """
+    transform = TextToTacotronSymbols(path, lang)
+    return transform(text)

modelscope/pydatasets/config.py

@@ -0,0 +1,22 @@
+import os
+from pathlib import Path
+
+# Cache location
+DEFAULT_CACHE_HOME = '~/.cache'
+CACHE_HOME = os.getenv('CACHE_HOME', DEFAULT_CACHE_HOME)
+DEFAULT_MS_CACHE_HOME = os.path.join(CACHE_HOME, 'modelscope/hub')
+MS_CACHE_HOME = os.path.expanduser(
+    os.getenv('MS_CACHE_HOME', DEFAULT_MS_CACHE_HOME))
+
+DEFAULT_MS_DATASETS_CACHE = os.path.join(MS_CACHE_HOME, 'datasets')
+MS_DATASETS_CACHE = Path(
+    os.getenv('MS_DATASETS_CACHE', DEFAULT_MS_DATASETS_CACHE))
+
+DOWNLOADED_DATASETS_DIR = 'downloads'
+DEFAULT_DOWNLOADED_DATASETS_PATH = os.path.join(MS_DATASETS_CACHE,
+                                                DOWNLOADED_DATASETS_DIR)
+DOWNLOADED_DATASETS_PATH = Path(
+    os.getenv('DOWNLOADED_DATASETS_PATH', DEFAULT_DOWNLOADED_DATASETS_PATH))
+
+MS_HUB_ENDPOINT = os.environ.get('MS_HUB_ENDPOINT',
+                                 'http://101.201.119.157:31752')

modelscope/pydatasets/py_dataset.py

@@ -1,64 +1,81 @@
-from typing import (Any, Callable, Dict, List, Mapping, Optional, Sequence,
-                    Union)
+import os
+from typing import (Any, Callable, Dict, Iterable, List, Mapping, Optional,
+                    Sequence, Union)
 
-from datasets import Dataset, load_dataset
+import numpy as np
+from datasets import Dataset
+from datasets import load_dataset as hf_load_dataset
+from datasets.config import TF_AVAILABLE, TORCH_AVAILABLE
+from datasets.packaged_modules import _PACKAGED_DATASETS_MODULES
+from datasets.utils.file_utils import (is_relative_path,
+                                       relative_to_absolute_path)
 
+from modelscope.pydatasets.config import MS_DATASETS_CACHE
+from modelscope.pydatasets.utils.ms_api import MsApi
 from modelscope.utils.constant import Hubs
 from modelscope.utils.logger import get_logger
 
 logger = get_logger()
 
 
+def format_list(para) -> List:
+    if para is None:
+        para = []
+    elif isinstance(para, str):
+        para = [para]
+    elif len(set(para)) < len(para):
+        raise ValueError(f'List columns contains duplicates: {para}')
+    return para
+
+
 class PyDataset:
     _hf_ds = None  # holds the underlying HuggingFace Dataset
     """A PyDataset backed by hugging face Dataset."""
 
-    def __init__(self, hf_ds: Dataset):
+    def __init__(self, hf_ds: Dataset, target: Optional[str] = None):
         self._hf_ds = hf_ds
-        self.target = None
+        self.target = target
 
     def __iter__(self):
-        if isinstance(self._hf_ds, Dataset):
-            for item in self._hf_ds:
-                if self.target is not None:
-                    yield item[self.target]
-                else:
-                    yield item
-        else:
-            for ds in self._hf_ds.values():
-                for item in ds:
-                    if self.target is not None:
-                        yield item[self.target]
-                    else:
-                        yield item
+        for item in self._hf_ds:
+            if self.target is not None:
+                yield item[self.target]
+            else:
+                yield item
+
+    def __getitem__(self, key):
+        return self._hf_ds[key]
 
     @classmethod
     def from_hf_dataset(cls,
                         hf_ds: Dataset,
-                        target: str = None) -> 'PyDataset':
-        dataset = cls(hf_ds)
-        dataset.target = target
-        return dataset
+                        target: str = None) -> Union[dict, 'PyDataset']:
+        if isinstance(hf_ds, Dataset):
+            return cls(hf_ds, target)
+        if len(hf_ds.keys()) == 1:
+            return cls(next(iter(hf_ds.values())), target)
+        return {k: cls(v, target) for k, v in hf_ds.items()}
 
     @staticmethod
-    def load(path: Union[str, list],
-             target: Optional[str] = None,
-             version: Optional[str] = None,
-             name: Optional[str] = None,
-             split: Optional[str] = None,
-             data_dir: Optional[str] = None,
-             data_files: Optional[Union[str, Sequence[str],
-                                        Mapping[str,
-                                                Union[str,
-                                                      Sequence[str]]]]] = None,
-             hub: Optional[Hubs] = None) -> 'PyDataset':
+    def load(
+        dataset_name: Union[str, list],
+        target: Optional[str] = None,
+        version: Optional[str] = None,
+        hub: Optional[Hubs] = Hubs.modelscope,
+        subset_name: Optional[str] = None,
+        split: Optional[str] = None,
+        data_dir: Optional[str] = None,
+        data_files: Optional[Union[str, Sequence[str],
+                                   Mapping[str, Union[str,
+                                                      Sequence[str]]]]] = None
+    ) -> Union[dict, 'PyDataset']:
         """Load a PyDataset from the ModelScope Hub, Hugging Face Hub, urls, or a local dataset.
             Args:
 
-                path (str): Path or name of the dataset.
+                dataset_name (str): Path or name of the dataset.
                 target (str, optional): Name of the column to output.
                 version (str, optional): Version of the dataset script to load:
-                name (str, optional): Defining the subset_name of the dataset.
+                subset_name (str, optional): Defining the subset_name of the dataset.
                 data_dir (str, optional): Defining the data_dir of the dataset configuration. I
                 data_files (str or Sequence or Mapping, optional): Path(s) to source data file(s).
                 split (str, optional): Which split of the data to load.
@@ -67,53 +84,302 @@ class PyDataset:
             Returns:
                 PyDataset (obj:`PyDataset`): PyDataset object for a certain dataset.
             """
-        if Hubs.modelscope == hub:
-            # TODO: parse data meta information from modelscope hub
-            # and possibly download data files to local (and update path)
-            print('getting data from modelscope hub')
-        if isinstance(path, str):
-            dataset = load_dataset(
-                path,
-                name=name,
+        if hub == Hubs.huggingface:
+            dataset = hf_load_dataset(
+                dataset_name,
+                name=subset_name,
                 revision=version,
                 split=split,
                 data_dir=data_dir,
                 data_files=data_files)
-        elif isinstance(path, list):
+            return PyDataset.from_hf_dataset(dataset, target=target)
+        else:
+            return PyDataset._load_ms_dataset(
+                dataset_name,
+                target=target,
+                subset_name=subset_name,
+                version=version,
+                split=split,
+                data_dir=data_dir,
+                data_files=data_files)
+
+    @staticmethod
+    def _load_ms_dataset(
+        dataset_name: Union[str, list],
+        target: Optional[str] = None,
+        version: Optional[str] = None,
+        subset_name: Optional[str] = None,
+        split: Optional[str] = None,
+        data_dir: Optional[str] = None,
+        data_files: Optional[Union[str, Sequence[str],
+                                   Mapping[str, Union[str,
+                                                      Sequence[str]]]]] = None
+    ) -> Union[dict, 'PyDataset']:
+        if isinstance(dataset_name, str):
+            use_hf = False
+            if dataset_name in _PACKAGED_DATASETS_MODULES or os.path.isdir(dataset_name) or \
+                    (os.path.isfile(dataset_name) and dataset_name.endswith('.py')):
+                use_hf = True
+            elif is_relative_path(dataset_name):
+                ms_api = MsApi()
+                dataset_scripts = ms_api.fetch_dataset_scripts(
+                    dataset_name, version)
+                if 'py' in dataset_scripts:  # dataset copied from hf datasets
+                    dataset_name = dataset_scripts['py'][0]
+                    use_hf = True
+            else:
+                raise FileNotFoundError(
+                    f"Couldn't find a dataset script at {relative_to_absolute_path(dataset_name)} "
+                    f'or any data file in the same directory.')
+
+            if use_hf:
+                dataset = hf_load_dataset(
+                    dataset_name,
+                    name=subset_name,
+                    revision=version,
+                    split=split,
+                    data_dir=data_dir,
+                    data_files=data_files,
+                    cache_dir=MS_DATASETS_CACHE)
+            else:
+                # TODO load from ms datahub
+                raise NotImplementedError(
+                    f'Dataset {dataset_name} load from modelscope datahub to be implemented in '
+                    f'the future')
+        elif isinstance(dataset_name, list):
             if target is None:
                 target = 'target'
-            dataset = Dataset.from_dict({target: [p] for p in path})
+            dataset = Dataset.from_dict({target: dataset_name})
         else:
             raise TypeError('path must be a str or a list, but got'
-                            f' {type(path)}')
+                            f' {type(dataset_name)}')
         return PyDataset.from_hf_dataset(dataset, target=target)
 
+    def to_torch_dataset_with_processors(
+        self,
+        preprocessors: Union[Callable, List[Callable]],
+        columns: Union[str, List[str]] = None,
+    ):
+        preprocessor_list = preprocessors if isinstance(
+            preprocessors, list) else [preprocessors]
+
+        columns = format_list(columns)
+
+        columns = [
+            key for key in self._hf_ds.features.keys() if key in columns
+        ]
+        sample = next(iter(self._hf_ds))
+
+        sample_res = {k: np.array(sample[k]) for k in columns}
+        for processor in preprocessor_list:
+            sample_res.update(
+                {k: np.array(v)
+                 for k, v in processor(sample).items()})
+
+        def is_numpy_number(value):
+            return np.issubdtype(value.dtype, np.integer) or np.issubdtype(
+                value.dtype, np.floating)
+
+        retained_columns = []
+        for k in sample_res.keys():
+            if not is_numpy_number(sample_res[k]):
+                logger.warning(
+                    f'Data of column {k} is non-numeric, will be removed')
+                continue
+            retained_columns.append(k)
+
+        import torch
+
+        class MsIterableDataset(torch.utils.data.IterableDataset):
+
+            def __init__(self, dataset: Iterable):
+                super(MsIterableDataset).__init__()
+                self.dataset = dataset
+
+            def __iter__(self):
+                for item_dict in self.dataset:
+                    res = {
+                        k: np.array(item_dict[k])
+                        for k in columns if k in retained_columns
+                    }
+                    for preprocessor in preprocessor_list:
+                        res.update({
+                            k: np.array(v)
+                            for k, v in preprocessor(item_dict).items()
+                            if k in retained_columns
+                        })
+                    yield res
+
+        return MsIterableDataset(self._hf_ds)
+
     def to_torch_dataset(
         self,
         columns: Union[str, List[str]] = None,
-        output_all_columns: bool = False,
+        preprocessors: Union[Callable, List[Callable]] = None,
         **format_kwargs,
     ):
-        self._hf_ds.reset_format()
-        self._hf_ds.set_format(
-            type='torch',
-            columns=columns,
-            output_all_columns=output_all_columns,
-            format_kwargs=format_kwargs)
-        return self._hf_ds
+        """Create a torch.utils.data.Dataset from the MS Dataset. The torch.utils.data.Dataset can be passed to
+           torch.utils.data.DataLoader.
+
+        Args:
+            preprocessors (Callable or List[Callable], default None): (list of) Preprocessor object used to process
+                every sample of the dataset. The output type of processors is dict, and each numeric field of the dict
+                will be used as a field of torch.utils.data.Dataset.
+            columns (str or List[str], default None): Dataset column(s) to be loaded (numeric data only). If the
+                preprocessor is None, the arg columns must have at least one column. If the `preprocessors` is not None,
+                the output fields of processors will also be added.
+            format_kwargs: A `dict` of arguments to be passed to the `torch.tensor`.
+
+        Returns:
+            :class:`tf.data.Dataset`
+
+        """
+        if not TORCH_AVAILABLE:
+            raise ImportError(
+                'The function to_torch_dataset requires pytorch to be installed'
+            )
+        if preprocessors is not None:
+            return self.to_torch_dataset_with_processors(preprocessors)
+        else:
+            self._hf_ds.reset_format()
+            self._hf_ds.set_format(
+                type='torch', columns=columns, format_kwargs=format_kwargs)
+            return self._hf_ds
+
+    def to_tf_dataset_with_processors(
+        self,
+        batch_size: int,
+        shuffle: bool,
+        preprocessors: Union[Callable, List[Callable]],
+        drop_remainder: bool = None,
+        prefetch: bool = True,
+        label_cols: Union[str, List[str]] = None,
+        columns: Union[str, List[str]] = None,
+    ):
+        preprocessor_list = preprocessors if isinstance(
+            preprocessors, list) else [preprocessors]
+
+        label_cols = format_list(label_cols)
+        columns = format_list(columns)
+        cols_to_retain = list(set(label_cols + columns))
+        retained_columns = [
+            key for key in self._hf_ds.features.keys() if key in cols_to_retain
+        ]
+        import tensorflow as tf
+        tf_dataset = tf.data.Dataset.from_tensor_slices(
+            np.arange(len(self._hf_ds), dtype=np.int64))
+        if shuffle:
+            tf_dataset = tf_dataset.shuffle(buffer_size=len(self._hf_ds))
+
+        def func(i, return_dict=False):
+            i = int(i)
+            res = {k: np.array(self._hf_ds[i][k]) for k in retained_columns}
+            for preprocessor in preprocessor_list:
+                # TODO preprocessor output may have the same key
+                res.update({
+                    k: np.array(v)
+                    for k, v in preprocessor(self._hf_ds[i]).items()
+                })
+            if return_dict:
+                return res
+            return tuple(list(res.values()))
+
+        sample_res = func(0, True)
+
+        @tf.function(input_signature=[tf.TensorSpec(None, tf.int64)])
+        def fetch_function(i):
+            output = tf.numpy_function(
+                func,
+                inp=[i],
+                Tout=[
+                    tf.dtypes.as_dtype(val.dtype)
+                    for val in sample_res.values()
+                ],
+            )
+            return {key: output[i] for i, key in enumerate(sample_res)}
+
+        tf_dataset = tf_dataset.map(
+            fetch_function, num_parallel_calls=tf.data.AUTOTUNE)
+        if label_cols:
+
+            def split_features_and_labels(input_batch):
+                labels = {
+                    key: tensor
+                    for key, tensor in input_batch.items() if key in label_cols
+                }
+                if len(input_batch) == 1:
+                    input_batch = next(iter(input_batch.values()))
+                if len(labels) == 1:
+                    labels = next(iter(labels.values()))
+                return input_batch, labels
+
+            tf_dataset = tf_dataset.map(split_features_and_labels)
+
+        elif len(columns) == 1:
+            tf_dataset = tf_dataset.map(lambda x: next(iter(x.values())))
+        if batch_size > 1:
+            tf_dataset = tf_dataset.batch(
+                batch_size, drop_remainder=drop_remainder)
+
+        if prefetch:
+            tf_dataset = tf_dataset.prefetch(tf.data.experimental.AUTOTUNE)
+        return tf_dataset
 
     def to_tf_dataset(
         self,
-        columns: Union[str, List[str]],
         batch_size: int,
         shuffle: bool,
-        collate_fn: Callable,
+        preprocessors: Union[Callable, List[Callable]] = None,
+        columns: Union[str, List[str]] = None,
+        collate_fn: Callable = None,
         drop_remainder: bool = None,
         collate_fn_args: Dict[str, Any] = None,
         label_cols: Union[str, List[str]] = None,
-        dummy_labels: bool = False,
         prefetch: bool = True,
     ):
+        """Create a tf.data.Dataset from the MS Dataset. This tf.data.Dataset can be passed to tf methods like
+           model.fit() or model.predict().
+
+        Args:
+            batch_size (int): Number of samples in a single batch.
+            shuffle(bool): Shuffle the dataset order.
+            preprocessors (Callable or List[Callable], default None): (list of) Preprocessor object used to process
+                every sample of the dataset. The output type of processors is dict, and each field of the dict will be
+                used as a field of the tf.data. Dataset. If the `preprocessors` is None, the `collate_fn`
+                shouldn't be None.
+            columns (str or List[str], default None): Dataset column(s) to be loaded. If the preprocessor is None,
+                the arg columns must have at least one column. If the `preprocessors` is not None, the output fields of
+                processors will also be added.
+            collate_fn(Callable, default None): A callable object used to collect lists of samples into a batch. If
+                the `preprocessors` is None, the `collate_fn` shouldn't be None.
+            drop_remainder(bool, default None): Drop the last incomplete batch when loading.
+            collate_fn_args (Dict, optional): A `dict` of arguments to be passed to the`collate_fn`.
+            label_cols (str or List[str], defalut None): Dataset column(s) to load as labels.
+            prefetch (bool, default True): Prefetch data.
+
+        Returns:
+            :class:`tf.data.Dataset`
+
+        """
+        if not TF_AVAILABLE:
+            raise ImportError(
+                'The function to_tf_dataset requires Tensorflow to be installed.'
+            )
+        if preprocessors is not None:
+            return self.to_tf_dataset_with_processors(
+                batch_size,
+                shuffle,
+                preprocessors,
+                drop_remainder=drop_remainder,
+                prefetch=prefetch,
+                label_cols=label_cols,
+                columns=columns)
+
+        if collate_fn is None:
+            logger.error(
+                'The `preprocessors` and the `collate_fn` should`t be both None.'
+            )
+            return None
         self._hf_ds.reset_format()
         return self._hf_ds.to_tf_dataset(
             columns,
@@ -123,7 +389,6 @@ class PyDataset:
             drop_remainder=drop_remainder,
             collate_fn_args=collate_fn_args,
             label_cols=label_cols,
-            dummy_labels=dummy_labels,
             prefetch=prefetch)
 
     def to_hf_dataset(self) -> Dataset:

modelscope/pydatasets/utils/ms_api.py

@@ -0,0 +1,66 @@
+import os
+from collections import defaultdict
+from typing import Optional
+
+import requests
+
+from modelscope.pydatasets.config import (DOWNLOADED_DATASETS_PATH,
+                                          MS_HUB_ENDPOINT)
+from modelscope.utils.logger import get_logger
+
+logger = get_logger()
+
+
+class MsApi:
+
+    def __init__(self, endpoint=MS_HUB_ENDPOINT):
+        self.endpoint = endpoint
+
+    def list_datasets(self):
+        path = f'{self.endpoint}/api/v1/datasets'
+        headers = None
+        params = {}
+        r = requests.get(path, params=params, headers=headers)
+        r.raise_for_status()
+        dataset_list = r.json()['Data']
+        return [x['Name'] for x in dataset_list]
+
+    def fetch_dataset_scripts(self,
+                              dataset_name: str,
+                              version: Optional[str] = 'master',
+                              force_download=False):
+        datahub_url = f'{self.endpoint}/api/v1/datasets?Query={dataset_name}'
+        r = requests.get(datahub_url)
+        r.raise_for_status()
+        dataset_list = r.json()['Data']
+        if len(dataset_list) == 0:
+            return None
+        dataset_id = dataset_list[0]['Id']
+        version = version or 'master'
+        datahub_url = f'{self.endpoint}/api/v1/datasets/{dataset_id}/repo/tree?Revision={version}'
+        r = requests.get(datahub_url)
+        r.raise_for_status()
+        file_list = r.json()['Data']['Files']
+        cache_dir = os.path.join(DOWNLOADED_DATASETS_PATH, dataset_name,
+                                 version)
+        os.makedirs(cache_dir, exist_ok=True)
+        local_paths = defaultdict(list)
+        for file_info in file_list:
+            file_path = file_info['Path']
+            if file_path.endswith('.py'):
+                datahub_url = f'{self.endpoint}/api/v1/datasets/{dataset_id}/repo/files?' \
+                              f'Revision={version}&Path={file_path}'
+                r = requests.get(datahub_url)
+                r.raise_for_status()
+                content = r.json()['Data']['Content']
+                local_path = os.path.join(cache_dir, file_path)
+                if os.path.exists(local_path) and not force_download:
+                    logger.warning(
+                        f"Reusing dataset {dataset_name}'s python file ({local_path})"
+                    )
+                    local_paths['py'].append(local_path)
+                    continue
+                with open(local_path, 'w') as f:
+                    f.writelines(content)
+                local_paths['py'].append(local_path)
+        return local_paths

modelscope/utils/audio/tts_exceptions.py

@@ -0,0 +1,42 @@
+"""
+Define TTS exceptions
+"""
+
+
+class TtsException(Exception):
+    """
+    TTS exception class.
+    """
+    pass
+
+
+class TtsFrontendException(TtsException):
+    """
+    TTS frontend module level exceptions.
+    """
+    pass
+
+
+class TtsFrontendInitializeFailedException(TtsFrontendException):
+    """
+    If tts frontend resource is invalid or not exist, this exception will be raised.
+    """
+    pass
+
+
+class TtsFrontendLanguageTypeInvalidException(TtsFrontendException):
+    """
+    If language type is invalid, this exception will be raised.
+    """
+
+
+class TtsVocoderException(TtsException):
+    """
+    Vocoder exception
+    """
+
+
+class TtsVocoderMelspecShapeMismatchException(TtsVocoderException):
+    """
+    If vocoder's input melspec shape mismatch, this exception will be raised.
+    """

modelscope/utils/constant.py

@@ -28,9 +28,11 @@ class Tasks(object):
     image_editing = 'image-editing'
     image_generation = 'image-generation'
     image_matting = 'image-matting'
+    ocr_detection = 'ocr-detection'
 
     # nlp tasks
     zero_shot_classification = 'zero-shot-classification'
+    word_segmentation = 'word-segmentation'
     sentiment_analysis = 'sentiment-analysis'
     sentence_similarity = 'sentence-similarity'
     text_classification = 'text-classification'

modelscope/utils/registry.py

@@ -67,7 +67,6 @@ class Registry(object):
         if module_name in self._modules[group_key]:
             raise KeyError(f'{module_name} is already registered in '
                            f'{self._name}[{group_key}]')
-
         self._modules[group_key][module_name] = module_cls
         module_cls.group_key = group_key
 

modelscope/utils/test_utils.py

@@ -2,6 +2,9 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
 import os
+import unittest
+
+from datasets.config import TF_AVAILABLE, TORCH_AVAILABLE
 
 TEST_LEVEL = 2
 TEST_LEVEL_STR = 'TEST_LEVEL'
@@ -15,6 +18,18 @@ def test_level():
     return TEST_LEVEL
 
 
+def require_tf(test_case):
+    if not TF_AVAILABLE:
+        test_case = unittest.skip('test requires TensorFlow')(test_case)
+    return test_case
+
+
+def require_torch(test_case):
+    if not TORCH_AVAILABLE:
+        test_case = unittest.skip('test requires PyTorch')(test_case)
+    return test_case
+
+
 def set_test_level(level: int):
     global TEST_LEVEL
     TEST_LEVEL = level

requirements.txt

@@ -2,4 +2,5 @@
 -r requirements/pipeline.txt
 -r requirements/multi-modal.txt
 -r requirements/nlp.txt
+-r requirements/audio.txt
 -r requirements/cv.txt

requirements/audio.txt

@@ -0,0 +1,26 @@
+#tts
+h5py==2.10.0
+#https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/TTS/requirements/ttsfrd-0.0.1-cp36-cp36m-linux_x86_64.whl
+https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/TTS/requirements/ttsfrd-0.0.1-cp37-cp37m-linux_x86_64.whl
+https://swap.oss-cn-hangzhou.aliyuncs.com/Jiaqi%2Fmaas%2Ftts%2Frequirements%2Fpytorch_wavelets-1.3.0-py3-none-any.whl?Expires=1685688388&OSSAccessKeyId=LTAI4Ffebq4d9jTVDwiSbY4L&Signature=jcQbg5EZ%2Bdys3%2F4BRn3srrKLdIg%3D
+#https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/TTS/requirements/ttsfrd-0.0.1-cp38-cp38-linux_x86_64.whl
+#https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/TTS/requirements/ttsfrd-0.0.1-cp39-cp39-linux_x86_64.whl
+inflect
+keras==2.2.4
+librosa
+lxml
+matplotlib
+nara_wpe
+numpy==1.18.*
+protobuf==3.20.*
+ptflops
+PyWavelets>=1.0.0
+scikit-learn==0.23.2
+sox
+tensorboard
+tensorflow==1.15.*
+torch==1.10.*
+torchaudio
+torchvision
+tqdm
+unidecode

requirements/cv.txt

@@ -1 +1,2 @@
 easydict
+tf_slim

requirements/nlp.txt

@@ -1 +1 @@
-https://alinlp.alibaba-inc.com/pypi/sofa-1.0.1.3-py3-none-any.whl
+https://alinlp.alibaba-inc.com/pypi/sofa-1.0.2-py3-none-any.whl

requirements/runtime.txt

@@ -1,12 +1,13 @@
 addict
 datasets
 easydict
-https://mindscope.oss-cn-hangzhou.aliyuncs.com/sdklib/maas_hub-0.2.2.dev0-py3-none-any.whl
+https://mindscope.oss-cn-hangzhou.aliyuncs.com/sdklib/maas_hub-0.2.4.dev0-py3-none-any.whl
 numpy
 opencv-python-headless
 Pillow>=6.2.0
 pyyaml
 requests
+scipy
 tokenizers<=0.10.3
 transformers<=4.16.2
 yapf

setup.cfg

@@ -11,6 +11,7 @@ default_section = THIRDPARTY
 BASED_ON_STYLE = pep8
 BLANK_LINE_BEFORE_NESTED_CLASS_OR_DEF = true
 SPLIT_BEFORE_EXPRESSION_AFTER_OPENING_PAREN = true
+SPLIT_BEFORE_ARITHMETIC_OPERATOR = true
 
 [codespell]
 skip = *.ipynb
@@ -20,5 +21,5 @@ ignore-words-list = patten,nd,ty,mot,hist,formating,winn,gool,datas,wan,confids
 [flake8]
 select = B,C,E,F,P,T4,W,B9
 max-line-length = 120
-ignore = F401,F821
+ignore = F401,F821,W503
 exclude = docs/src,*.pyi,.git