diff --git a/data/test/images/image_camouflag_detection.jpg b/data/test/images/image_camouflag_detection.jpg deleted file mode 100644 index 5029067d..00000000 --- a/data/test/images/image_camouflag_detection.jpg +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:4c713215f7fb4da5382c9137347ee52956a7a44d5979c4cffd3c9b6d1d7e878f -size 19445 diff --git a/modelscope/models/cv/salient_detection/models/__init__.py b/modelscope/models/cv/salient_detection/models/__init__.py index 6df5101a..8ea7a5d3 100644 --- a/modelscope/models/cv/salient_detection/models/__init__.py +++ b/modelscope/models/cv/salient_detection/models/__init__.py @@ -1,4 +1,3 @@ # The implementation is adopted from U-2-Net, made publicly available under the Apache 2.0 License # source code avaiable via https://github.com/xuebinqin/U-2-Net -from .senet import SENet from .u2net import U2NET diff --git a/modelscope/models/cv/salient_detection/models/modules.py b/modelscope/models/cv/salient_detection/models/modules.py deleted file mode 100644 index 09796bd3..00000000 --- a/modelscope/models/cv/salient_detection/models/modules.py +++ /dev/null @@ -1,178 +0,0 @@ -# Copyright (c) Alibaba, Inc. and its affiliates. -import torch -import torch.nn as nn -import torch.nn.functional as F - -from .utils import ConvBNReLU - - -class AreaLayer(nn.Module): - - def __init__(self, in_channel, out_channel): - super(AreaLayer, self).__init__() - self.lbody = nn.Sequential( - nn.Conv2d(out_channel, out_channel, 1), - nn.BatchNorm2d(out_channel), nn.ReLU(inplace=True)) - self.hbody = nn.Sequential( - nn.Conv2d(in_channel, out_channel, 1), nn.BatchNorm2d(out_channel), - nn.ReLU(inplace=True)) - self.body = nn.Sequential( - nn.Conv2d(2 * out_channel, out_channel, 3, 1, 1), - nn.BatchNorm2d(out_channel), nn.ReLU(inplace=True), - nn.Conv2d(out_channel, out_channel, 3, 1, 1), - nn.BatchNorm2d(out_channel), nn.ReLU(inplace=True), - nn.Conv2d(out_channel, 1, 1)) - - def forward(self, xl, xh): - xl1 = self.lbody(xl) - xl1 = F.interpolate( - xl1, size=xh.size()[2:], mode='bilinear', align_corners=True) - xh1 = self.hbody(xh) - x = torch.cat((xl1, xh1), dim=1) - x_out = self.body(x) - return x_out - - -class EdgeLayer(nn.Module): - - def __init__(self, in_channel, out_channel): - super(EdgeLayer, self).__init__() - self.lbody = nn.Sequential( - nn.Conv2d(out_channel, out_channel, 1), - nn.BatchNorm2d(out_channel), nn.ReLU(inplace=True)) - self.hbody = nn.Sequential( - nn.Conv2d(in_channel, out_channel, 1), nn.BatchNorm2d(out_channel), - nn.ReLU(inplace=True)) - self.bodye = nn.Sequential( - nn.Conv2d(2 * out_channel, out_channel, 3, 1, 1), - nn.BatchNorm2d(out_channel), nn.ReLU(inplace=True), - nn.Conv2d(out_channel, out_channel, 3, 1, 1), - nn.BatchNorm2d(out_channel), nn.ReLU(inplace=True), - nn.Conv2d(out_channel, 1, 1)) - - def forward(self, xl, xh): - xl1 = self.lbody(xl) - xh1 = self.hbody(xh) - xh1 = F.interpolate( - xh1, size=xl.size()[2:], mode='bilinear', align_corners=True) - x = torch.cat((xl1, xh1), dim=1) - x_out = self.bodye(x) - return x_out - - -class EBlock(nn.Module): - - def __init__(self, inchs, outchs): - super(EBlock, self).__init__() - self.elayer = nn.Sequential( - ConvBNReLU(inchs + 1, outchs, kernel_size=3, padding=1, stride=1), - ConvBNReLU(outchs, outchs, 1)) - self.salayer = nn.Sequential( - nn.Conv2d(2, 1, 3, 1, 1, bias=False), - nn.BatchNorm2d(1, momentum=0.01), nn.Sigmoid()) - - def forward(self, x, edgeAtten): - x = torch.cat((x, edgeAtten), dim=1) - ex = self.elayer(x) - ex_max = torch.max(ex, 1, keepdim=True)[0] - ex_mean = torch.mean(ex, dim=1, keepdim=True) - xei_compress = torch.cat((ex_max, ex_mean), dim=1) - - scale = self.salayer(xei_compress) - x_out = ex * scale - return x_out - - -class StructureE(nn.Module): - - def __init__(self, inchs, outchs, EM): - super(StructureE, self).__init__() - self.ne_modules = int(inchs / EM) - NM = int(outchs / self.ne_modules) - elayes = [] - for i in range(self.ne_modules): - emblock = EBlock(EM, NM) - elayes.append(emblock) - self.emlayes = nn.ModuleList(elayes) - self.body = nn.Sequential( - ConvBNReLU(outchs, outchs, 3, 1, 1), ConvBNReLU(outchs, outchs, 1)) - - def forward(self, x, edgeAtten): - if edgeAtten.size() != x.size(): - edgeAtten = F.interpolate( - edgeAtten, x.size()[2:], mode='bilinear', align_corners=False) - xx = torch.chunk(x, self.ne_modules, dim=1) - efeas = [] - for i in range(self.ne_modules): - xei = self.emlayes[i](xx[i], edgeAtten) - efeas.append(xei) - efeas = torch.cat(efeas, dim=1) - x_out = self.body(efeas) - return x_out - - -class ABlock(nn.Module): - - def __init__(self, inchs, outchs, k): - super(ABlock, self).__init__() - self.alayer = nn.Sequential( - ConvBNReLU(inchs, outchs, k, 1, k // 2), - ConvBNReLU(outchs, outchs, 1)) - self.arlayer = nn.Sequential( - ConvBNReLU(inchs, outchs, k, 1, k // 2), - ConvBNReLU(outchs, outchs, 1)) - self.fusion = ConvBNReLU(2 * outchs, outchs, 1) - - def forward(self, x, areaAtten): - xa = x * areaAtten - xra = x * (1 - areaAtten) - xout = self.fusion(torch.cat((xa, xra), dim=1)) - return xout - - -class AMFusion(nn.Module): - - def __init__(self, inchs, outchs, AM): - super(AMFusion, self).__init__() - self.k = [3, 3, 5, 5] - self.conv_up = ConvBNReLU(inchs, outchs, 3, 1, 1) - self.up = nn.Upsample( - scale_factor=2, mode='bilinear', align_corners=True) - self.na_modules = int(outchs / AM) - alayers = [] - for i in range(self.na_modules): - layer = ABlock(AM, AM, self.k[i]) - alayers.append(layer) - self.alayers = nn.ModuleList(alayers) - self.fusion_0 = ConvBNReLU(outchs, outchs, 3, 1, 1) - self.fusion_e = nn.Sequential( - nn.Conv2d( - outchs, outchs, kernel_size=(3, 1), padding=(1, 0), - bias=False), nn.BatchNorm2d(outchs), nn.ReLU(inplace=True), - nn.Conv2d( - outchs, outchs, kernel_size=(1, 3), padding=(0, 1), - bias=False), nn.BatchNorm2d(outchs), nn.ReLU(inplace=True)) - self.fusion_e1 = nn.Sequential( - nn.Conv2d( - outchs, outchs, kernel_size=(5, 1), padding=(2, 0), - bias=False), nn.BatchNorm2d(outchs), nn.ReLU(inplace=True), - nn.Conv2d( - outchs, outchs, kernel_size=(1, 5), padding=(0, 2), - bias=False), nn.BatchNorm2d(outchs), nn.ReLU(inplace=True)) - self.fusion = ConvBNReLU(3 * outchs, outchs, 1) - - def forward(self, xl, xh, xhm): - xh1 = self.up(self.conv_up(xh)) - x = xh1 + xl - xm = self.up(torch.sigmoid(xhm)) - xx = torch.chunk(x, self.na_modules, dim=1) - xxmids = [] - for i in range(self.na_modules): - xi = self.alayers[i](xx[i], xm) - xxmids.append(xi) - xfea = torch.cat(xxmids, dim=1) - x0 = self.fusion_0(xfea) - x1 = self.fusion_e(xfea) - x2 = self.fusion_e1(xfea) - x_out = self.fusion(torch.cat((x0, x1, x2), dim=1)) - return x_out diff --git a/modelscope/models/cv/salient_detection/models/senet.py b/modelscope/models/cv/salient_detection/models/senet.py deleted file mode 100644 index 37cf42be..00000000 --- a/modelscope/models/cv/salient_detection/models/senet.py +++ /dev/null @@ -1,74 +0,0 @@ -# Copyright (c) Alibaba, Inc. and its affiliates. -import torch -import torch.nn as nn -import torch.nn.functional as F - -from .backbone import res2net50_v1b_26w_4s as res2net -from .modules import AMFusion, AreaLayer, EdgeLayer, StructureE -from .utils import ASPP, CBAM, ConvBNReLU - - -class SENet(nn.Module): - - def __init__(self, backbone_path=None, pretrained=False): - super(SENet, self).__init__() - resnet50 = res2net(backbone_path, pretrained) - self.layer0_1 = nn.Sequential(resnet50.conv1, resnet50.bn1, - resnet50.relu) - self.maxpool = resnet50.maxpool - self.layer1 = resnet50.layer1 - self.layer2 = resnet50.layer2 - self.layer3 = resnet50.layer3 - self.layer4 = resnet50.layer4 - self.aspp3 = ASPP(1024, 256) - self.aspp4 = ASPP(2048, 256) - self.cbblock3 = CBAM(inchs=256, kernel_size=5) - self.cbblock4 = CBAM(inchs=256, kernel_size=5) - self.up = nn.Upsample( - mode='bilinear', scale_factor=2, align_corners=False) - self.conv_up = ConvBNReLU(512, 512, 1) - self.aux_edge = EdgeLayer(512, 256) - self.aux_area = AreaLayer(512, 256) - self.layer1_enhance = StructureE(256, 128, 128) - self.layer2_enhance = StructureE(512, 256, 128) - self.layer3_decoder = AMFusion(512, 256, 128) - self.layer2_decoder = AMFusion(256, 128, 128) - self.out_conv_8 = nn.Conv2d(256, 1, 1) - self.out_conv_4 = nn.Conv2d(128, 1, 1) - - def forward(self, x): - layer0 = self.layer0_1(x) - layer0s = self.maxpool(layer0) - layer1 = self.layer1(layer0s) - layer2 = self.layer2(layer1) - layer3 = self.layer3(layer2) - layer4 = self.layer4(layer3) - layer3_eh = self.cbblock3(self.aspp3(layer3)) - layer4_eh = self.cbblock4(self.aspp4(layer4)) - layer34 = self.conv_up( - torch.cat((self.up(layer4_eh), layer3_eh), dim=1)) - edge_atten = self.aux_edge(layer1, layer34) - area_atten = self.aux_area(layer1, layer34) - edge_atten_ = torch.sigmoid(edge_atten) - layer1_eh = self.layer1_enhance(layer1, edge_atten_) - layer2_eh = self.layer2_enhance(layer2, edge_atten_) - layer2_fu = self.layer3_decoder(layer2_eh, layer34, area_atten) - out_8 = self.out_conv_8(layer2_fu) - layer1_fu = self.layer2_decoder(layer1_eh, layer2_fu, out_8) - out_4 = self.out_conv_4(layer1_fu) - out_16 = F.interpolate( - area_atten, - size=x.size()[2:], - mode='bilinear', - align_corners=False) - out_8 = F.interpolate( - out_8, size=x.size()[2:], mode='bilinear', align_corners=False) - out_4 = F.interpolate( - out_4, size=x.size()[2:], mode='bilinear', align_corners=False) - edge_out = F.interpolate( - edge_atten_, - size=x.size()[2:], - mode='bilinear', - align_corners=False) - - return out_4.sigmoid(), out_8.sigmoid(), out_16.sigmoid(), edge_out diff --git a/modelscope/models/cv/salient_detection/salient_model.py b/modelscope/models/cv/salient_detection/salient_model.py index e25166c8..73c3c3fb 100644 --- a/modelscope/models/cv/salient_detection/salient_model.py +++ b/modelscope/models/cv/salient_detection/salient_model.py @@ -2,6 +2,7 @@ import os.path as osp import cv2 +import numpy as np import torch from PIL import Image from torchvision import transforms @@ -9,9 +10,8 @@ from torchvision import transforms from modelscope.metainfo import Models from modelscope.models.base.base_torch_model import TorchModel from modelscope.models.builder import MODELS -from modelscope.utils.config import Config from modelscope.utils.constant import ModelFile, Tasks -from .models import U2NET, SENet +from .models import U2NET @MODELS.register_module( @@ -22,25 +22,13 @@ class SalientDetection(TorchModel): """str -- model file root.""" super().__init__(model_dir, *args, **kwargs) model_path = osp.join(model_dir, ModelFile.TORCH_MODEL_FILE) - - self.norm_mean = [0.485, 0.456, 0.406] - self.norm_std = [0.229, 0.224, 0.225] - self.norm_size = (320, 320) - - config_path = osp.join(model_dir, 'config.py') - if osp.exists(config_path) is False: - self.model = U2NET(3, 1) - else: - self.model = SENet(backbone_path=None, pretrained=False) - config = Config.from_file(config_path) - self.norm_mean = config.norm_mean - self.norm_std = config.norm_std - self.norm_size = config.norm_size + self.model = U2NET(3, 1) checkpoint = torch.load(model_path, map_location='cpu') self.transform_input = transforms.Compose([ - transforms.Resize(self.norm_size), + transforms.Resize((320, 320)), transforms.ToTensor(), - transforms.Normalize(mean=self.norm_mean, std=self.norm_std) + transforms.Normalize( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) self.model.load_state_dict(checkpoint) self.model.eval() diff --git a/modelscope/pipelines/cv/image_salient_detection_pipeline.py b/modelscope/pipelines/cv/image_salient_detection_pipeline.py index 4b4df52c..4a3eaa65 100644 --- a/modelscope/pipelines/cv/image_salient_detection_pipeline.py +++ b/modelscope/pipelines/cv/image_salient_detection_pipeline.py @@ -12,11 +12,6 @@ from modelscope.utils.constant import Tasks @PIPELINES.register_module( Tasks.semantic_segmentation, module_name=Pipelines.salient_detection) -@PIPELINES.register_module( - Tasks.semantic_segmentation, - module_name=Pipelines.salient_boudary_detection) -@PIPELINES.register_module( - Tasks.semantic_segmentation, module_name=Pipelines.camouflaged_detection) class ImageSalientDetectionPipeline(Pipeline): def __init__(self, model: str, **kwargs): diff --git a/tests/pipelines/test_salient_detection.py b/tests/pipelines/test_salient_detection.py index 3101213c..bcb904e6 100644 --- a/tests/pipelines/test_salient_detection.py +++ b/tests/pipelines/test_salient_detection.py @@ -23,27 +23,6 @@ class SalientDetectionTest(unittest.TestCase, DemoCompatibilityCheck): import cv2 cv2.imwrite(input_location + '_salient.jpg', result[OutputKeys.MASKS]) - @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_salient_boudary_detection(self): - input_location = 'data/test/images/image_salient_detection.jpg' - model_id = 'damo/cv_res2net_salient-detection' - salient_detect = pipeline(Tasks.semantic_segmentation, model=model_id) - result = salient_detect(input_location) - import cv2 - cv2.imwrite(input_location + '_boudary_salient.jpg', - result[OutputKeys.MASKS]) - - @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_camouflag_detection(self): - input_location = 'data/test/images/image_camouflag_detection.jpg' - model_id = 'damo/cv_res2net_camouflaged-detection' - camouflag_detect = pipeline( - Tasks.semantic_segmentation, model=model_id) - result = camouflag_detect(input_location) - import cv2 - cv2.imwrite(input_location + '_camouflag.jpg', - result[OutputKeys.MASKS]) - @unittest.skip('demo compatibility test is only enabled on a needed-basis') def test_demo_compatibility(self): self.compatibility_check()