add syncdreamer as a image-to-3d pipeline (#679)

* add syncdreamer

* delete demo_output

* remove useless files and hide clip install

* update pipeline

* update clip package

* update unitest

modelscope/metainfo.py

@@ -127,6 +127,7 @@ class Models(object):
     human_image_generation = 'human-image-generation'
     image_view_transform = 'image-view-transform'
     image_control_3d_portrait = 'image-control-3d-portrait'
+    
 
     # nlp models
     bert = 'bert'
@@ -455,6 +456,7 @@ class Pipelines(object):
     human3d_animation = 'human3d-animation'
     image_view_transform = 'image-view-transform'
     image_control_3d_portrait = 'image-control-3d-portrait'
+    image_to_3d = 'image-to-3d'
 
     # nlp tasks
     automatic_post_editing = 'automatic-post-editing'

modelscope/models/cv/image_to_3d/__init__.py

@@ -0,0 +1,2 @@
+# Copyright 2021-2022 The Alibaba Fundamental Vision Team Authors. All rights reserved.
+from . import ldm

modelscope/models/cv/image_to_3d/ldm/base_utils.py

@@ -0,0 +1,158 @@
+import pickle
+import numpy as np
+import cv2
+from skimage.io import imread
+
+
+def save_pickle(data, pkl_path):
+    # os.system('mkdir -p {}'.format(os.path.dirname(pkl_path)))
+    with open(pkl_path, 'wb') as f:
+        pickle.dump(data, f)
+
+def read_pickle(pkl_path):
+    with open(pkl_path, 'rb') as f:
+        return pickle.load(f)
+
+def draw_epipolar_line(F, img0, img1, pt0, color):
+    h1,w1=img1.shape[:2]
+    hpt = np.asarray([pt0[0], pt0[1], 1], dtype=np.float32)[:, None]
+    l = F @ hpt
+    l = l[:, 0]
+    a, b, c = l[0], l[1], l[2]
+    pt1 = np.asarray([0, -c / b]).astype(np.int32)
+    pt2 = np.asarray([w1, (-a * w1 - c) / b]).astype(np.int32)
+
+    img0 = cv2.circle(img0, tuple(pt0.astype(np.int32)), 5, color, 2)
+    img1 = cv2.line(img1, tuple(pt1), tuple(pt2), color, 2)
+    return img0, img1
+
+def draw_epipolar_lines(F, img0, img1,num=20):
+    img0,img1=img0.copy(),img1.copy()
+    h0, w0, _ = img0.shape
+    h1, w1, _ = img1.shape
+
+    for k in range(num):
+        color = np.random.randint(0, 255, [3], dtype=np.int32)
+        color = [int(c) for c in color]
+        pt = np.random.uniform(0, 1, 2)
+        pt[0] *= w0
+        pt[1] *= h0
+        pt = pt.astype(np.int32)
+        img0, img1 = draw_epipolar_line(F, img0, img1, pt, color)
+
+    return img0, img1
+
+def compute_F(K1, K2, Rt0, Rt1=None):
+    if Rt1 is None:
+        R, t = Rt0[:,:3], Rt0[:,3:]
+    else:
+        Rt = compute_dR_dt(Rt0,Rt1)
+        R, t = Rt[:,:3], Rt[:,3:]
+    A = K1 @ R.T @ t # [3,1]
+    C = np.asarray([[0,-A[2,0],A[1,0]],
+                    [A[2,0],0,-A[0,0]],
+                    [-A[1,0],A[0,0],0]])
+    F = (np.linalg.inv(K2)).T @ R @ K1.T @ C
+    return F
+
+def compute_dR_dt(Rt0, Rt1):
+    R0, t0 = Rt0[:,:3], Rt0[:,3:]
+    R1, t1 = Rt1[:,:3], Rt1[:,3:]
+    dR = np.dot(R1, R0.T)
+    dt = t1 - np.dot(dR, t0)
+    return np.concatenate([dR, dt], -1)
+
+def concat_images(img0,img1,vert=False):
+    if not vert:
+        h0,h1=img0.shape[0],img1.shape[0],
+        if h0<h1: img0=cv2.copyMakeBorder(img0,0,h1-h0,0,0,borderType=cv2.BORDER_CONSTANT,value=0)
+        if h1<h0: img1=cv2.copyMakeBorder(img1,0,h0-h1,0,0,borderType=cv2.BORDER_CONSTANT,value=0)
+        img = np.concatenate([img0, img1], axis=1)
+    else:
+        w0,w1=img0.shape[1],img1.shape[1]
+        if w0<w1: img0=cv2.copyMakeBorder(img0,0,0,0,w1-w0,borderType=cv2.BORDER_CONSTANT,value=0)
+        if w1<w0: img1=cv2.copyMakeBorder(img1,0,0,0,w0-w1,borderType=cv2.BORDER_CONSTANT,value=0)
+        img = np.concatenate([img0, img1], axis=0)
+
+    return img
+
+def concat_images_list(*args,vert=False):
+    if len(args)==1: return args[0]
+    img_out=args[0]
+    for img in args[1:]:
+        img_out=concat_images(img_out,img,vert)
+    return img_out
+
+
+def pose_inverse(pose):
+    R = pose[:,:3].T
+    t = - R @ pose[:,3:]
+    return np.concatenate([R,t],-1)
+
+def project_points(pts,RT,K):
+    pts = np.matmul(pts,RT[:,:3].transpose())+RT[:,3:].transpose()
+    pts = np.matmul(pts,K.transpose())
+    dpt = pts[:,2]
+    mask0 = (np.abs(dpt)<1e-4) & (np.abs(dpt)>0)
+    if np.sum(mask0)>0: dpt[mask0]=1e-4
+    mask1=(np.abs(dpt) > -1e-4) & (np.abs(dpt) < 0)
+    if np.sum(mask1)>0: dpt[mask1]=-1e-4
+    pts2d = pts[:,:2]/dpt[:,None]
+    return pts2d, dpt
+
+
+def draw_keypoints(img, kps, colors=None, radius=2):
+    out_img=img.copy()
+    for pi, pt in enumerate(kps):
+        pt = np.round(pt).astype(np.int32)
+        if colors is not None:
+            color=[int(c) for c in colors[pi]]
+            cv2.circle(out_img, tuple(pt), radius, color, -1)
+        else:
+            cv2.circle(out_img, tuple(pt), radius, (0,255,0), -1)
+    return out_img
+
+
+def output_points(fn,pts,colors=None):
+    with open(fn, 'w') as f:
+        for pi, pt in enumerate(pts):
+            f.write(f'{pt[0]:.6f} {pt[1]:.6f} {pt[2]:.6f} ')
+            if colors is not None:
+                f.write(f'{int(colors[pi,0])} {int(colors[pi,1])} {int(colors[pi,2])}')
+            f.write('\n')
+
+DEPTH_MAX, DEPTH_MIN = 2.4, 0.6
+DEPTH_VALID_MAX, DEPTH_VALID_MIN = 2.37, 0.63
+def read_depth_objaverse(depth_fn):
+    depth = imread(depth_fn)
+    depth = depth.astype(np.float32) / 65535 * (DEPTH_MAX-DEPTH_MIN) + DEPTH_MIN
+    mask = (depth > DEPTH_VALID_MIN) & (depth < DEPTH_VALID_MAX)
+    return depth, mask
+
+
+def mask_depth_to_pts(mask,depth,K,rgb=None):
+    hs,ws=np.nonzero(mask)
+    depth=depth[hs,ws]
+    pts=np.asarray([ws,hs,depth],np.float32).transpose()
+    pts[:,:2]*=pts[:,2:]
+    if rgb is not None:
+        return np.dot(pts, np.linalg.inv(K).transpose()), rgb[hs,ws]
+    else:
+        return np.dot(pts, np.linalg.inv(K).transpose())
+
+def transform_points_pose(pts, pose):
+    R, t = pose[:, :3], pose[:, 3]
+    if len(pts.shape)==1:
+        return (R @ pts[:,None] + t[:,None])[:,0]
+    return pts @ R.T + t[None,:]
+
+def pose_apply(pose,pts):
+    return transform_points_pose(pts, pose)
+
+def downsample_gaussian_blur(img, ratio):
+    sigma = (1 / ratio) / 3
+    # ksize=np.ceil(2*sigma)
+    ksize = int(np.ceil(((sigma - 0.8) / 0.3 + 1) * 2 + 1))
+    ksize = ksize + 1 if ksize % 2 == 0 else ksize
+    img = cv2.GaussianBlur(img, (ksize, ksize), sigma, borderType=cv2.BORDER_REFLECT101)
+    return img

modelscope/models/cv/image_to_3d/ldm/models/autoencoder.py

@@ -0,0 +1,443 @@
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+
+from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
+
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.model import Encoder, Decoder
+from modelscope.models.cv.image_to_3d.ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from modelscope.models.cv.image_to_3d.ldm.util import instantiate_from_config
+
+
+class VQModel(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 n_embed,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 batch_resize_range=None,
+                 scheduler_config=None,
+                 lr_g_factor=1.0,
+                 remap=None,
+                 sane_index_shape=False, # tell vector quantizer to return indices as bhw
+                 use_ema=False
+                 ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.n_embed = n_embed
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
+                                        remap=remap,
+                                        sane_index_shape=sane_index_shape)
+        self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        self.batch_resize_range = batch_resize_range
+        if self.batch_resize_range is not None:
+            print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        self.scheduler_config = scheduler_config
+        self.lr_g_factor = lr_g_factor
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info
+
+    def encode_to_prequant(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, quant):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+    def decode_code(self, code_b):
+        quant_b = self.quantize.embed_code(code_b)
+        dec = self.decode(quant_b)
+        return dec
+
+    def forward(self, input, return_pred_indices=False):
+        quant, diff, (_,_,ind) = self.encode(input)
+        dec = self.decode(quant)
+        if return_pred_indices:
+            return dec, diff, ind
+        return dec, diff
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        if self.batch_resize_range is not None:
+            lower_size = self.batch_resize_range[0]
+            upper_size = self.batch_resize_range[1]
+            if self.global_step <= 4:
+                # do the first few batches with max size to avoid later oom
+                new_resize = upper_size
+            else:
+                new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
+            if new_resize != x.shape[2]:
+                x = F.interpolate(x, size=new_resize, mode="bicubic")
+            x = x.detach()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        # https://github.com/pytorch/pytorch/issues/37142
+        # try not to fool the heuristics
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train",
+                                            predicted_indices=ind)
+
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, suffix=""):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
+                                        self.global_step,
+                                        last_layer=self.get_last_layer(),
+                                        split="val"+suffix,
+                                        predicted_indices=ind
+                                        )
+
+        discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
+                                            self.global_step,
+                                            last_layer=self.get_last_layer(),
+                                            split="val"+suffix,
+                                            predicted_indices=ind
+                                            )
+        rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log(f"val{suffix}/rec_loss", rec_loss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        self.log(f"val{suffix}/aeloss", aeloss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        if version.parse(pl.__version__) >= version.parse('1.4.0'):
+            del log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr_d = self.learning_rate
+        lr_g = self.lr_g_factor*self.learning_rate
+        print("lr_d", lr_d)
+        print("lr_g", lr_g)
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quantize.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr_g, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr_d, betas=(0.5, 0.9))
+
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+                {
+                    'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+            ]
+            return [opt_ae, opt_disc], scheduler
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if only_inputs:
+            log["inputs"] = x
+            return log
+        xrec, _ = self(x)
+        if x.shape[1] > 3:
+            # colorize with random projection
+            assert xrec.shape[1] > 3
+            x = self.to_rgb(x)
+            xrec = self.to_rgb(xrec)
+        log["inputs"] = x
+        log["reconstructions"] = xrec
+        if plot_ema:
+            with self.ema_scope():
+                xrec_ema, _ = self(x)
+                if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
+                log["reconstructions_ema"] = xrec_ema
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class VQModelInterface(VQModel):
+    def __init__(self, embed_dim, *args, **kwargs):
+        super().__init__(embed_dim=embed_dim, *args, **kwargs)
+        self.embed_dim = embed_dim
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path}")
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x

modelscope/models/cv/image_to_3d/ldm/models/diffusion/sync_dreamer.py

@@ -0,0 +1,673 @@
+from pathlib import Path
+
+import pytorch_lightning as pl
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from skimage.io import imsave
+from torch.optim.lr_scheduler import LambdaLR
+from tqdm import tqdm
+
+from modelscope.models.cv.image_to_3d.ldm.base_utils import read_pickle, concat_images_list
+from modelscope.models.cv.image_to_3d.ldm.models.diffusion.sync_dreamer_utils import get_warp_coordinates, create_target_volume
+from modelscope.models.cv.image_to_3d.ldm.models.diffusion.sync_dreamer_network import NoisyTargetViewEncoder, SpatialTime3DNet, FrustumTV3DNet
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.util import make_ddim_timesteps, timestep_embedding
+from modelscope.models.cv.image_to_3d.ldm.modules.encoders.modules import FrozenCLIPImageEmbedder
+from modelscope.models.cv.image_to_3d.ldm.util import instantiate_from_config
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+def disable_training_module(module: nn.Module):
+    module = module.eval()
+    module.train = disabled_train
+    for para in module.parameters():
+        para.requires_grad = False
+    return module
+
+def repeat_to_batch(tensor, B, VN):
+    t_shape = tensor.shape
+    ones = [1 for _ in range(len(t_shape)-1)]
+    tensor_new = tensor.view(B,1,*t_shape[1:]).repeat(1,VN,*ones).view(B*VN,*t_shape[1:])
+    return tensor_new
+
+class UNetWrapper(nn.Module):
+    def __init__(self, diff_model_config, drop_conditions=False, drop_scheme='default', use_zero_123=True):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.drop_conditions = drop_conditions
+        self.drop_scheme=drop_scheme
+        self.use_zero_123 = use_zero_123
+
+
+    def drop(self, cond, mask):
+        shape = cond.shape
+        B = shape[0]
+        cond = mask.view(B,*[1 for _ in range(len(shape)-1)]) * cond
+        return cond
+
+    def get_trainable_parameters(self):
+        return self.diffusion_model.get_trainable_parameters()
+
+    def get_drop_scheme(self, B, device):
+        if self.drop_scheme=='default':
+            random = torch.rand(B, dtype=torch.float32, device=device)
+            drop_clip = (random > 0.15) & (random <= 0.2)
+            drop_volume = (random > 0.1) & (random <= 0.15)
+            drop_concat = (random > 0.05) & (random <= 0.1)
+            drop_all = random <= 0.05
+        else:
+            raise NotImplementedError
+        return drop_clip, drop_volume, drop_concat, drop_all
+
+    def forward(self, x, t, clip_embed, volume_feats, x_concat, is_train=False):
+        """
+
+        @param x:             B,4,H,W
+        @param t:             B,
+        @param clip_embed:    B,M,768
+        @param volume_feats:  B,C,D,H,W
+        @param x_concat:      B,C,H,W
+        @param is_train:
+        @return:
+        """
+        if self.drop_conditions and is_train:
+            B = x.shape[0]
+            drop_clip, drop_volume, drop_concat, drop_all = self.get_drop_scheme(B, x.device)
+
+            clip_mask = 1.0 - (drop_clip | drop_all).float()
+            clip_embed = self.drop(clip_embed, clip_mask)
+
+            volume_mask = 1.0 - (drop_volume | drop_all).float()
+            for k, v in volume_feats.items():
+                volume_feats[k] = self.drop(v, mask=volume_mask)
+
+            concat_mask = 1.0 - (drop_concat | drop_all).float()
+            x_concat = self.drop(x_concat, concat_mask)
+
+        if self.use_zero_123:
+            # zero123 does not multiply this when encoding, maybe a bug for zero123
+            first_stage_scale_factor = 0.18215
+            x_concat_ = x_concat * 1.0
+            x_concat_[:, :4] = x_concat_[:, :4] / first_stage_scale_factor
+        else:
+            x_concat_ = x_concat
+
+        x = torch.cat([x, x_concat_], 1)
+        pred = self.diffusion_model(x, t, clip_embed, source_dict=volume_feats)
+        return pred
+
+    def predict_with_unconditional_scale(self, x, t, clip_embed, volume_feats, x_concat, unconditional_scale):
+        x_ = torch.cat([x] * 2, 0)
+        t_ = torch.cat([t] * 2, 0)
+        clip_embed_ = torch.cat([clip_embed, torch.zeros_like(clip_embed)], 0)
+
+        v_ = {}
+        for k, v in volume_feats.items():
+            v_[k] = torch.cat([v, torch.zeros_like(v)], 0)
+
+        x_concat_ = torch.cat([x_concat, torch.zeros_like(x_concat)], 0)
+        if self.use_zero_123:
+            # zero123 does not multiply this when encoding, maybe a bug for zero123
+            first_stage_scale_factor = 0.18215
+            x_concat_[:, :4] = x_concat_[:, :4] / first_stage_scale_factor
+        x_ = torch.cat([x_, x_concat_], 1)
+        s, s_uc = self.diffusion_model(x_, t_, clip_embed_, source_dict=v_).chunk(2)
+        s = s_uc + unconditional_scale * (s - s_uc)
+        return s
+
+
+class SpatialVolumeNet(nn.Module):
+    def __init__(self, time_dim, view_dim, view_num,
+                 input_image_size=256, frustum_volume_depth=48,
+                 spatial_volume_size=32, spatial_volume_length=0.5,
+                 frustum_volume_length=0.86603 # sqrt(3)/2
+                 ):
+        super().__init__()
+        self.target_encoder = NoisyTargetViewEncoder(time_dim, view_dim, output_dim=16)
+        self.spatial_volume_feats = SpatialTime3DNet(input_dim=16 * view_num, time_dim=time_dim, dims=(64, 128, 256, 512))
+        self.frustum_volume_feats = FrustumTV3DNet(64, time_dim, view_dim, dims=(64, 128, 256, 512))
+
+        self.frustum_volume_length = frustum_volume_length
+        self.input_image_size = input_image_size
+        self.spatial_volume_size = spatial_volume_size
+        self.spatial_volume_length = spatial_volume_length
+
+        self.frustum_volume_size = self.input_image_size // 8
+        self.frustum_volume_depth = frustum_volume_depth
+        self.time_dim = time_dim
+        self.view_dim = view_dim
+        self.default_origin_depth = 1.5 # our rendered images are 1.5 away from the origin, we assume camera is 1.5 away from the origin
+
+    def construct_spatial_volume(self, x, t_embed, v_embed, target_poses, target_Ks):
+        """
+        @param x:            B,N,4,H,W
+        @param t_embed:      B,t_dim
+        @param v_embed:      B,N,v_dim
+        @param target_poses: N,3,4
+        @param target_Ks:    N,3,3
+        @return:
+        """
+        B, N, _, H, W = x.shape
+        V = self.spatial_volume_size
+        device = x.device
+
+        spatial_volume_verts = torch.linspace(-self.spatial_volume_length, self.spatial_volume_length, V, dtype=torch.float32, device=device)
+        spatial_volume_verts = torch.stack(torch.meshgrid(spatial_volume_verts, spatial_volume_verts, spatial_volume_verts), -1)
+        spatial_volume_verts = spatial_volume_verts.reshape(1, V ** 3, 3)[:, :, (2, 1, 0)]
+        spatial_volume_verts = spatial_volume_verts.view(1, V, V, V, 3).permute(0, 4, 1, 2, 3).repeat(B, 1, 1, 1, 1)
+
+        # encode source features
+        t_embed_ = t_embed.view(B, 1, self.time_dim).repeat(1, N, 1).view(B, N, self.time_dim)
+        # v_embed_ = v_embed.view(1, N, self.view_dim).repeat(B, 1, 1).view(B, N, self.view_dim)
+        v_embed_ = v_embed
+        target_Ks = target_Ks.unsqueeze(0).repeat(B, 1, 1, 1)
+        target_poses = target_poses.unsqueeze(0).repeat(B, 1, 1, 1)
+
+        # extract 2D image features
+        spatial_volume_feats = []
+        # project source features
+        for ni in range(0, N):
+            pose_source_ = target_poses[:, ni]
+            K_source_ = target_Ks[:, ni]
+            x_ = self.target_encoder(x[:, ni], t_embed_[:, ni], v_embed_[:, ni])
+            C = x_.shape[1]
+
+            coords_source = get_warp_coordinates(spatial_volume_verts, x_.shape[-1], self.input_image_size, K_source_, pose_source_).view(B, V, V * V, 2)
+            unproj_feats_ = F.grid_sample(x_, coords_source, mode='bilinear', padding_mode='zeros', align_corners=True)
+            unproj_feats_ = unproj_feats_.view(B, C, V, V, V)
+            spatial_volume_feats.append(unproj_feats_)
+
+        spatial_volume_feats = torch.stack(spatial_volume_feats, 1) # B,N,C,V,V,V
+        N = spatial_volume_feats.shape[1]
+        spatial_volume_feats = spatial_volume_feats.view(B, N*C, V, V, V)
+
+        spatial_volume_feats = self.spatial_volume_feats(spatial_volume_feats, t_embed)  # b,64,32,32,32
+        return spatial_volume_feats
+
+    def construct_view_frustum_volume(self, spatial_volume, t_embed, v_embed, poses, Ks, target_indices):
+        """
+        @param spatial_volume:    B,C,V,V,V
+        @param t_embed:           B,t_dim
+        @param v_embed:           B,N,v_dim
+        @param poses:             N,3,4
+        @param Ks:                N,3,3
+        @param target_indices:    B,TN
+        @return: B*TN,C,H,W
+        """
+        B, TN = target_indices.shape
+        H, W = self.frustum_volume_size, self.frustum_volume_size
+        D = self.frustum_volume_depth
+        V = self.spatial_volume_size
+
+        near = torch.ones(B * TN, 1, H, W, dtype=spatial_volume.dtype, device=spatial_volume.device) * self.default_origin_depth - self.frustum_volume_length
+        far = torch.ones(B * TN, 1, H, W, dtype=spatial_volume.dtype, device=spatial_volume.device) * self.default_origin_depth + self.frustum_volume_length
+
+        target_indices = target_indices.view(B*TN) # B*TN
+        poses_ = poses[target_indices] # B*TN,3,4
+        Ks_ = Ks[target_indices] # B*TN,3,4
+        volume_xyz, volume_depth = create_target_volume(D, self.frustum_volume_size, self.input_image_size, poses_, Ks_, near, far) # B*TN,3 or 1,D,H,W
+
+        volume_xyz_ = volume_xyz / self.spatial_volume_length  # since the spatial volume is constructed in [-spatial_volume_length,spatial_volume_length]
+        volume_xyz_ = volume_xyz_.permute(0, 2, 3, 4, 1)  # B*TN,D,H,W,3
+        spatial_volume_ = spatial_volume.unsqueeze(1).repeat(1, TN, 1, 1, 1, 1).view(B * TN, -1, V, V, V)
+        volume_feats = F.grid_sample(spatial_volume_, volume_xyz_, mode='bilinear', padding_mode='zeros', align_corners=True) # B*TN,C,D,H,W
+
+        v_embed_ = v_embed[torch.arange(B)[:,None], target_indices.view(B,TN)].view(B*TN, -1) # B*TN
+        t_embed_ = t_embed.unsqueeze(1).repeat(1,TN,1).view(B*TN,-1)
+        volume_feats_dict = self.frustum_volume_feats(volume_feats, t_embed_, v_embed_)
+        return volume_feats_dict, volume_depth
+"""
+    SyncDreamer is a SoTA Novel View Synthesis model which can generate 16 consistent views seamlessly.
+    Please refer to: https://arxiv.org/abs/2309.03453 for more technique details.
+"""
+class SyncMultiviewDiffusion(pl.LightningModule):
+    def __init__(self, unet_config, scheduler_config,
+                 finetune_unet=False, finetune_projection=True,
+                 view_num=16, image_size=256,
+                 cfg_scale=3.0, output_num=8, batch_view_num=4,
+                 drop_conditions=False, drop_scheme='default',
+                 clip_image_encoder_path="/apdcephfs/private_rondyliu/projects/clip/ViT-L-14.pt"):
+        super().__init__()
+
+        self.finetune_unet = finetune_unet
+        self.finetune_projection = finetune_projection
+
+        self.view_num = view_num
+        self.viewpoint_dim = 4
+        self.output_num = output_num
+        self.image_size = image_size
+
+        self.batch_view_num = batch_view_num
+        self.cfg_scale = cfg_scale
+
+        self.clip_image_encoder_path = clip_image_encoder_path
+
+        self._init_time_step_embedding()
+        self._init_first_stage()
+        self._init_schedule()
+        self._init_multiview()
+        self._init_clip_image_encoder()
+        self._init_clip_projection()
+
+        self.spatial_volume = SpatialVolumeNet(self.time_embed_dim, self.viewpoint_dim, self.view_num)
+        self.model = UNetWrapper(unet_config, drop_conditions=drop_conditions, drop_scheme=drop_scheme)
+        self.scheduler_config = scheduler_config
+
+        latent_size = image_size//8
+        self.ddim = SyncDDIMSampler(self, 200, "uniform", 1.0, latent_size=latent_size)
+
+    def _init_clip_projection(self):
+        self.cc_projection = nn.Linear(772, 768)
+        nn.init.eye_(list(self.cc_projection.parameters())[0][:768, :768])
+        nn.init.zeros_(list(self.cc_projection.parameters())[1])
+        self.cc_projection.requires_grad_(True)
+
+        if not self.finetune_projection:
+            disable_training_module(self.cc_projection)
+
+    def _init_multiview(self):
+        K, azs, _, _, poses = read_pickle(self.clip_image_encoder_path.replace("ViT-L-14.pt",f'camera-{self.view_num}.pkl'))
+        default_image_size = 256
+        ratio = self.image_size/default_image_size
+        K = np.diag([ratio,ratio,1]) @ K
+        K = torch.from_numpy(K.astype(np.float32)) # [3,3]
+        K = K.unsqueeze(0).repeat(self.view_num,1,1)        # N,3,3
+        poses = torch.from_numpy(poses.astype(np.float32))  # N,3,4
+        self.register_buffer('poses', poses)
+        self.register_buffer('Ks', K)
+        azs = (azs + np.pi) % (np.pi * 2) - np.pi # scale to [-pi,pi] and the index=0 has az=0
+        self.register_buffer('azimuth', torch.from_numpy(azs.astype(np.float32)))
+
+    def get_viewpoint_embedding(self, batch_size, elevation_ref):
+        """
+        @param batch_size:
+        @param elevation_ref: B
+        @return:
+        """
+        azimuth_input = self.azimuth[0].unsqueeze(0) # 1
+        azimuth_target = self.azimuth # N
+        elevation_input = -elevation_ref # note that zero123 use a negative elevation here!!!
+        elevation_target = -np.deg2rad(30)
+        d_e = elevation_target - elevation_input # B
+        N = self.azimuth.shape[0]
+        B = batch_size
+        d_e = d_e.unsqueeze(1).repeat(1, N)
+        d_a = azimuth_target - azimuth_input # N
+        d_a = d_a.unsqueeze(0).repeat(B, 1)
+        d_z = torch.zeros_like(d_a)
+        embedding = torch.stack([d_e, torch.sin(d_a), torch.cos(d_a), d_z], -1) # B,N,4
+        return embedding
+
+    def _init_first_stage(self):
+        first_stage_config={
+            "target": "modelscope.models.cv.image_to_3d.ldm.models.autoencoder.AutoencoderKL",
+            "params": {
+                "embed_dim": 4,
+                "monitor": "val/rec_loss",
+                "ddconfig":{
+                  "double_z": True,
+                  "z_channels": 4,
+                  "resolution": self.image_size,
+                  "in_channels": 3,
+                  "out_ch": 3,
+                  "ch": 128,
+                  "ch_mult": [1,2,4,4],
+                  "num_res_blocks": 2,
+                  "attn_resolutions": [],
+                  "dropout": 0.0
+                },
+                "lossconfig": {"target": "torch.nn.Identity"},
+            }
+        }
+        self.first_stage_scale_factor = 0.18215
+        self.first_stage_model = instantiate_from_config(first_stage_config)
+        self.first_stage_model = disable_training_module(self.first_stage_model)
+
+    def _init_clip_image_encoder(self):
+        self.clip_image_encoder = FrozenCLIPImageEmbedder(model=self.clip_image_encoder_path)
+        self.clip_image_encoder = disable_training_module(self.clip_image_encoder)
+
+    def _init_schedule(self):
+        self.num_timesteps = 1000
+        linear_start = 0.00085
+        linear_end = 0.0120
+        num_timesteps = 1000
+        betas = torch.linspace(linear_start ** 0.5, linear_end ** 0.5, num_timesteps, dtype=torch.float32) ** 2 # T
+        assert betas.shape[0] == self.num_timesteps
+
+        # all in float64 first
+        alphas = 1. - betas
+        alphas_cumprod = torch.cumprod(alphas, dim=0) # T
+        alphas_cumprod_prev = torch.cat([torch.ones(1, dtype=torch.float64), alphas_cumprod[:-1]], 0)
+        posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod) # T
+        posterior_log_variance_clipped = torch.log(torch.clamp(posterior_variance, min=1e-20))
+        posterior_log_variance_clipped = torch.clamp(posterior_log_variance_clipped, min=-10)
+
+        self.register_buffer("betas", betas.float())
+        self.register_buffer("alphas", alphas.float())
+        self.register_buffer("alphas_cumprod", alphas_cumprod.float())
+        self.register_buffer("sqrt_alphas_cumprod", torch.sqrt(alphas_cumprod).float())
+        self.register_buffer("sqrt_one_minus_alphas_cumprod", torch.sqrt(1 - alphas_cumprod).float())
+        self.register_buffer("posterior_variance", posterior_variance.float())
+        self.register_buffer('posterior_log_variance_clipped', posterior_log_variance_clipped.float())
+
+    def _init_time_step_embedding(self):
+        self.time_embed_dim = 256
+        self.time_embed = nn.Sequential(
+            nn.Linear(self.time_embed_dim, self.time_embed_dim),
+            nn.SiLU(True),
+            nn.Linear(self.time_embed_dim, self.time_embed_dim),
+        )
+
+    def encode_first_stage(self, x, sample=True):
+        with torch.no_grad():
+            posterior = self.first_stage_model.encode(x)  # b,4,h//8,w//8
+            if sample:
+                return posterior.sample().detach() * self.first_stage_scale_factor
+            else:
+                return posterior.mode().detach() * self.first_stage_scale_factor
+
+    def decode_first_stage(self, z):
+        with torch.no_grad():
+            z = 1. / self.first_stage_scale_factor * z
+            return self.first_stage_model.decode(z)
+
+    def prepare(self, batch):
+        # encode target
+        if 'target_image' in batch:
+            image_target = batch['target_image'].permute(0, 1, 4, 2, 3) # b,n,3,h,w
+            N = image_target.shape[1]
+            x = [self.encode_first_stage(image_target[:,ni], True) for ni in range(N)]
+            x = torch.stack(x, 1) # b,n,4,h//8,w//8
+        else:
+            x = None
+
+        image_input = batch['input_image'].permute(0, 3, 1, 2)
+        elevation_input = batch['input_elevation'][:, 0] # b
+        x_input = self.encode_first_stage(image_input)
+        input_info = {'image': image_input, 'elevation': elevation_input, 'x': x_input}
+        with torch.no_grad():
+            clip_embed = self.clip_image_encoder.encode(image_input)
+        return x, clip_embed, input_info
+
+    def embed_time(self, t):
+        t_embed = timestep_embedding(t, self.time_embed_dim, repeat_only=False) # B,TED
+        t_embed = self.time_embed(t_embed) # B,TED
+        return t_embed
+
+    def get_target_view_feats(self, x_input, spatial_volume, clip_embed, t_embed, v_embed, target_index):
+        """
+        @param x_input:        B,4,H,W
+        @param spatial_volume: B,C,V,V,V
+        @param clip_embed:     B,1,768
+        @param t_embed:        B,t_dim
+        @param v_embed:        B,N,v_dim
+        @param target_index:   B,TN
+        @return:
+            tensors of size B*TN,*
+        """
+        B, _, H, W = x_input.shape
+        frustum_volume_feats, frustum_volume_depth = self.spatial_volume.construct_view_frustum_volume(spatial_volume, t_embed, v_embed, self.poses, self.Ks, target_index)
+
+        # clip
+        TN = target_index.shape[1]
+        v_embed_ = v_embed[torch.arange(B)[:,None], target_index].view(B*TN, self.viewpoint_dim) # B*TN,v_dim
+        clip_embed_ = clip_embed.unsqueeze(1).repeat(1,TN,1,1).view(B*TN,1,768)
+        clip_embed_ = self.cc_projection(torch.cat([clip_embed_, v_embed_.unsqueeze(1)], -1))  # B*TN,1,768
+
+        x_input_ = x_input.unsqueeze(1).repeat(1, TN, 1, 1, 1).view(B * TN, 4, H, W)
+
+        x_concat = x_input_
+        return clip_embed_, frustum_volume_feats, x_concat
+
+    def training_step(self, batch):
+        B = batch['image'].shape[0]
+        time_steps = torch.randint(0, self.num_timesteps, (B,), device=self.device).long()
+
+        x, clip_embed, input_info = self.prepare(batch)
+        x_noisy, noise = self.add_noise(x, time_steps)  # B,N,4,H,W
+
+        N = self.view_num
+        target_index = torch.randint(0, N, (B, 1), device=self.device).long() # B, 1
+        v_embed = self.get_viewpoint_embedding(B, input_info['elevation']) # N,v_dim
+
+        t_embed = self.embed_time(time_steps)
+        spatial_volume = self.spatial_volume.construct_spatial_volume(x_noisy, t_embed, v_embed, self.poses, self.Ks)
+
+        clip_embed, volume_feats, x_concat = self.get_target_view_feats(input_info['x'], spatial_volume, clip_embed, t_embed, v_embed, target_index)
+
+        x_noisy_ = x_noisy[torch.arange(B)[:,None],target_index][:,0] # B,4,H,W
+        noise_predict = self.model(x_noisy_, time_steps, clip_embed, volume_feats, x_concat, is_train=True) # B,4,H,W
+
+        noise_target = noise[torch.arange(B)[:,None],target_index][:,0] # B,4,H,W
+        # loss simple for diffusion
+        loss_simple = torch.nn.functional.mse_loss(noise_target, noise_predict, reduction='none')
+        loss = loss_simple.mean()
+        self.log('sim', loss_simple.mean(), prog_bar=True, logger=True, on_step=True, on_epoch=True, rank_zero_only=True)
+
+        # log others
+        lr = self.optimizers().param_groups[0]['lr']
+        self.log('lr', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False, rank_zero_only=True)
+        self.log("step", self.global_step, prog_bar=True, logger=True, on_step=True, on_epoch=False, rank_zero_only=True)
+        return loss
+
+    def add_noise(self, x_start, t):
+        """
+        @param x_start: B,*
+        @param t:       B,
+        @return:
+        """
+        B = x_start.shape[0]
+        noise = torch.randn_like(x_start) # B,*
+
+        sqrt_alphas_cumprod_  = self.sqrt_alphas_cumprod[t] # B,
+        sqrt_one_minus_alphas_cumprod_ = self.sqrt_one_minus_alphas_cumprod[t] # B
+        sqrt_alphas_cumprod_ = sqrt_alphas_cumprod_.view(B, *[1 for _ in range(len(x_start.shape)-1)])
+        sqrt_one_minus_alphas_cumprod_ = sqrt_one_minus_alphas_cumprod_.view(B, *[1 for _ in range(len(x_start.shape)-1)])
+        x_noisy = sqrt_alphas_cumprod_ * x_start + sqrt_one_minus_alphas_cumprod_ * noise
+        return x_noisy, noise
+
+    def sample(self, batch, cfg_scale, batch_view_num, use_ddim=True,
+               return_inter_results=False, inter_interval=50, inter_view_interval=2):
+        _, clip_embed, input_info = self.prepare(batch)
+        if use_ddim:
+            x_sample, inter = self.ddim.sample(input_info, clip_embed, unconditional_scale=cfg_scale, log_every_t=inter_interval, batch_view_num=batch_view_num)
+        else:
+            raise NotImplementedError
+
+        N = x_sample.shape[1]
+        x_sample = torch.stack([self.decode_first_stage(x_sample[:, ni]) for ni in range(N)], 1)
+        if return_inter_results:
+            torch.cuda.synchronize()
+            torch.cuda.empty_cache()
+            inter = torch.stack(inter['x_inter'], 2) # # B,N,T,C,H,W
+            B,N,T,C,H,W = inter.shape
+            inter_results = []
+            for ni in tqdm(range(0, N, inter_view_interval)):
+                inter_results_ = []
+                for ti in range(T):
+                    inter_results_.append(self.decode_first_stage(inter[:, ni, ti]))
+                inter_results.append(torch.stack(inter_results_, 1)) # B,T,3,H,W
+            inter_results = torch.stack(inter_results,1) # B,N,T,3,H,W
+            return x_sample, inter_results
+        else:
+            return x_sample
+
+    def log_image(self,  x_sample, batch, step, output_dir, only_first_row=False):
+        process = lambda x: ((torch.clip(x, min=-1, max=1).cpu().numpy() * 0.5 + 0.5) * 255).astype(np.uint8)
+        B = x_sample.shape[0]
+        N = x_sample.shape[1]
+        image_cond = []
+        for bi in range(B):
+            img_pr_ = concat_images_list(process(batch['ref_image'][bi]),*[process(x_sample[bi, ni].permute(1, 2, 0)) for ni in range(N)])
+            img_gt_ = concat_images_list(process(batch['ref_image'][bi]),*[process(batch['image'][bi, ni]) for ni in range(N)])
+            if not only_first_row or bi==0:
+                image_cond.append(concat_images_list(img_gt_, img_pr_, vert=True))
+            else:
+                image_cond.append(img_pr_)
+
+
+        output_dir = Path(output_dir)
+        imsave(str(output_dir/f'{step}.jpg'), concat_images_list(*image_cond, vert=True))
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        if batch_idx==0 and self.global_rank==0:
+            self.eval()
+            step = self.global_step
+            batch_ = {}
+            for k, v in batch.items(): batch_[k] = v[:self.output_num]
+            x_sample = self.sample(batch_, self.cfg_scale, self.batch_view_num)
+            output_dir = Path(self.image_dir) / 'images' / 'val'
+            output_dir.mkdir(exist_ok=True, parents=True)
+            self.log_image(x_sample, batch, step, output_dir=output_dir)
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        print(f'setting learning rate to {lr:.4f} ...')
+        paras = []
+        if self.finetune_projection:
+            paras.append({"params": self.cc_projection.parameters(), "lr": lr},)
+        if self.finetune_unet:
+            paras.append({"params": self.model.parameters(), "lr": lr},)
+        else:
+            paras.append({"params": self.model.get_trainable_parameters(), "lr": lr},)
+
+        paras.append({"params": self.time_embed.parameters(), "lr": lr*10.0},)
+        paras.append({"params": self.spatial_volume.parameters(), "lr": lr*10.0},)
+
+        opt = torch.optim.AdamW(paras, lr=lr)
+
+        scheduler = instantiate_from_config(self.scheduler_config)
+        print("Setting up LambdaLR scheduler...")
+        scheduler = [{'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule), 'interval': 'step', 'frequency': 1}]
+        return [opt], scheduler
+
+class SyncDDIMSampler:
+    def __init__(self, model: SyncMultiviewDiffusion, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., latent_size=32):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.latent_size = latent_size
+        self._make_schedule(ddim_num_steps, ddim_discretize, ddim_eta)
+        self.eta = ddim_eta
+
+    def _make_schedule(self,  ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, num_ddpm_timesteps=self.ddpm_num_timesteps, verbose=verbose) # DT
+        ddim_timesteps_ = torch.from_numpy(self.ddim_timesteps.astype(np.int64)) # DT
+
+        alphas_cumprod = self.model.alphas_cumprod # T
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        self.ddim_alphas = alphas_cumprod[ddim_timesteps_].double() # DT
+        self.ddim_alphas_prev = torch.cat([alphas_cumprod[0:1], alphas_cumprod[ddim_timesteps_[:-1]]], 0) # DT
+        self.ddim_sigmas = ddim_eta * torch.sqrt((1 - self.ddim_alphas_prev) / (1 - self.ddim_alphas) * (1 - self.ddim_alphas / self.ddim_alphas_prev))
+
+        self.ddim_alphas_raw = self.model.alphas[ddim_timesteps_].float() # DT
+        self.ddim_sigmas = self.ddim_sigmas.float()
+        self.ddim_alphas = self.ddim_alphas.float()
+        self.ddim_alphas_prev = self.ddim_alphas_prev.float()
+        self.ddim_sqrt_one_minus_alphas = torch.sqrt(1. - self.ddim_alphas).float()
+
+
+    @torch.no_grad()
+    def denoise_apply_impl(self, x_target_noisy, index, noise_pred, is_step0=False):
+        """
+        @param x_target_noisy: B,N,4,H,W
+        @param index:          index
+        @param noise_pred:     B,N,4,H,W
+        @param is_step0:       bool
+        @return:
+        """
+        device = x_target_noisy.device
+        B,N,_,H,W = x_target_noisy.shape
+
+        # apply noise
+        a_t = self.ddim_alphas[index].to(device).float().view(1,1,1,1,1)
+        a_prev = self.ddim_alphas_prev[index].to(device).float().view(1,1,1,1,1)
+        sqrt_one_minus_at = self.ddim_sqrt_one_minus_alphas[index].to(device).float().view(1,1,1,1,1)
+        sigma_t = self.ddim_sigmas[index].to(device).float().view(1,1,1,1,1)
+
+        pred_x0 = (x_target_noisy - sqrt_one_minus_at * noise_pred) / a_t.sqrt()
+        dir_xt = torch.clamp(1. - a_prev - sigma_t**2, min=1e-7).sqrt() * noise_pred
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt
+        if not is_step0:
+            noise = sigma_t * torch.randn_like(x_target_noisy)
+            x_prev = x_prev + noise
+        return x_prev
+
+    @torch.no_grad()
+    def denoise_apply(self, x_target_noisy, input_info, clip_embed, time_steps, index, unconditional_scale, batch_view_num=1, is_step0=False):
+        """
+        @param x_target_noisy:   B,N,4,H,W
+        @param input_info:
+        @param clip_embed:       B,M,768
+        @param time_steps:       B,
+        @param index:            int
+        @param unconditional_scale:
+        @param batch_view_num:   int
+        @param is_step0:         bool
+        @return:
+        """
+        x_input, elevation_input = input_info['x'], input_info['elevation']
+        B, N, C, H, W = x_target_noisy.shape
+
+        # construct source data
+        v_embed = self.model.get_viewpoint_embedding(B, elevation_input) # B,N,v_dim
+        t_embed = self.model.embed_time(time_steps)  # B,t_dim
+        spatial_volume = self.model.spatial_volume.construct_spatial_volume(x_target_noisy, t_embed, v_embed, self.model.poses, self.model.Ks)
+
+        e_t = []
+        target_indices = torch.arange(N) # N
+        for ni in range(0, N, batch_view_num):
+            x_target_noisy_ = x_target_noisy[:, ni:ni + batch_view_num]
+            VN = x_target_noisy_.shape[1]
+            x_target_noisy_ = x_target_noisy_.reshape(B*VN,C,H,W)
+
+            time_steps_ = repeat_to_batch(time_steps, B, VN)
+            target_indices_ = target_indices[ni:ni+batch_view_num].unsqueeze(0).repeat(B,1)
+            clip_embed_, volume_feats_, x_concat_ = self.model.get_target_view_feats(x_input, spatial_volume, clip_embed, t_embed, v_embed, target_indices_)
+            if unconditional_scale!=1.0:
+                noise = self.model.model.predict_with_unconditional_scale(x_target_noisy_, time_steps_, clip_embed_, volume_feats_, x_concat_, unconditional_scale)
+            else:
+                noise = self.model.model(x_target_noisy_, time_steps_, clip_embed_, volume_feats_, x_concat_, is_train=False)
+            e_t.append(noise.view(B,VN,4,H,W))
+
+        e_t = torch.cat(e_t, 1)
+        x_prev = self.denoise_apply_impl(x_target_noisy, index, e_t, is_step0)
+        return x_prev
+
+    @torch.no_grad()
+    def sample(self, input_info, clip_embed, unconditional_scale=1.0, log_every_t=50, batch_view_num=1):
+        """
+        @param input_info:      x, elevation
+        @param clip_embed:      B,M,768
+        @param unconditional_scale:
+        @param log_every_t:
+        @param batch_view_num:
+        @return:
+        """
+        print(f"unconditional scale {unconditional_scale:.1f}")
+        C, H, W = 4, self.latent_size, self.latent_size
+        B = clip_embed.shape[0]
+        N = self.model.view_num
+        device = self.model.device
+        x_target_noisy = torch.randn([B, N, C, H, W], device=device)
+
+        timesteps = self.ddim_timesteps
+        intermediates = {'x_inter': []}
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+
+        iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1 # index in ddim state
+            time_steps = torch.full((B,), step, device=device, dtype=torch.long)
+            x_target_noisy = self.denoise_apply(x_target_noisy, input_info, clip_embed, time_steps, index, unconditional_scale, batch_view_num=batch_view_num, is_step0=index==0)
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(x_target_noisy)
+
+        return x_target_noisy, intermediates

modelscope/models/cv/image_to_3d/ldm/models/diffusion/sync_dreamer_attention.py

@@ -0,0 +1,142 @@
+import torch
+import torch.nn as nn
+
+from modelscope.models.cv.image_to_3d.ldm.modules.attention import default, zero_module, checkpoint
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.openaimodel import UNetModel
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.util import timestep_embedding
+
+class DepthAttention(nn.Module):
+    def __init__(self, query_dim, context_dim, heads, dim_head, output_bias=True):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = nn.Conv2d(query_dim, inner_dim, 1, 1, bias=False)
+        self.to_k = nn.Conv3d(context_dim, inner_dim, 1, 1, bias=False)
+        self.to_v = nn.Conv3d(context_dim, inner_dim, 1, 1, bias=False)
+        if output_bias:
+            self.to_out = nn.Conv2d(inner_dim, query_dim, 1, 1)
+        else:
+            self.to_out = nn.Conv2d(inner_dim, query_dim, 1, 1, bias=False)
+
+    def forward(self, x, context):
+        """
+
+        @param x:        b,f0,h,w
+        @param context:  b,f1,d,h,w
+        @return:
+        """
+        hn, hd = self.heads, self.dim_head
+        b, _, h, w = x.shape
+        b, _, d, h, w = context.shape
+
+        q = self.to_q(x).reshape(b,hn,hd,h,w) # b,t,h,w
+        k = self.to_k(context).reshape(b,hn,hd,d,h,w) # b,t,d,h,w
+        v = self.to_v(context).reshape(b,hn,hd,d,h,w) # b,t,d,h,w
+
+        sim = torch.sum(q.unsqueeze(3) * k, 2) * self.scale # b,hn,d,h,w
+        attn = sim.softmax(dim=2)
+
+        # b,hn,hd,d,h,w * b,hn,1,d,h,w
+        out = torch.sum(v * attn.unsqueeze(2), 3) # b,hn,hd,h,w
+        out = out.reshape(b,hn*hd,h,w)
+        return self.to_out(out)
+
+
+class DepthTransformer(nn.Module):
+    def __init__(self, dim, n_heads, d_head, context_dim=None, checkpoint=True):
+        super().__init__()
+        inner_dim = n_heads * d_head
+        self.proj_in = nn.Sequential(
+            nn.Conv2d(dim, inner_dim, 1, 1),
+            nn.GroupNorm(8, inner_dim),
+            nn.SiLU(True),
+        )
+        self.proj_context = nn.Sequential(
+            nn.Conv3d(context_dim, context_dim, 1, 1, bias=False), # no bias
+            nn.GroupNorm(8, context_dim),
+            nn.ReLU(True), # only relu, because we want input is 0, output is 0
+        )
+        self.depth_attn = DepthAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, context_dim=context_dim, output_bias=False)  # is a self-attention if not self.disable_self_attn
+        self.proj_out = nn.Sequential(
+            nn.GroupNorm(8, inner_dim),
+            nn.ReLU(True),
+            nn.Conv2d(inner_dim, inner_dim, 3, 1, 1, bias=False),
+            nn.GroupNorm(8, inner_dim),
+            nn.ReLU(True),
+            zero_module(nn.Conv2d(inner_dim, dim, 3, 1, 1, bias=False)),
+        )
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context):
+        x_in = x
+        x = self.proj_in(x)
+        context = self.proj_context(context)
+        x = self.depth_attn(x, context)
+        x = self.proj_out(x) + x_in
+        return x
+
+
+class DepthWiseAttention(UNetModel):
+    def __init__(self, volume_dims=(5,16,32,64), *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # num_heads = 4
+        model_channels = kwargs['model_channels']
+        channel_mult = kwargs['channel_mult']
+        d0,d1,d2,d3 = volume_dims
+
+        # 4
+        ch = model_channels*channel_mult[2]
+        self.middle_conditions = DepthTransformer(ch, 4, d3 // 2, context_dim=d3)
+
+        self.output_conditions=nn.ModuleList()
+        self.output_b2c = {3:0,4:1,5:2,6:3,7:4,8:5,9:6,10:7,11:8}
+        # 8
+        ch = model_channels*channel_mult[2]
+        self.output_conditions.append(DepthTransformer(ch, 4, d2 // 2, context_dim=d2)) # 0
+        self.output_conditions.append(DepthTransformer(ch, 4, d2 // 2, context_dim=d2)) # 1
+        # 16
+        self.output_conditions.append(DepthTransformer(ch, 4, d1 // 2, context_dim=d1)) # 2
+        ch = model_channels*channel_mult[1]
+        self.output_conditions.append(DepthTransformer(ch, 4, d1 // 2, context_dim=d1)) # 3
+        self.output_conditions.append(DepthTransformer(ch, 4, d1 // 2, context_dim=d1)) # 4
+        # 32
+        self.output_conditions.append(DepthTransformer(ch, 4, d0 // 2, context_dim=d0)) # 5
+        ch = model_channels*channel_mult[0]
+        self.output_conditions.append(DepthTransformer(ch, 4, d0 // 2, context_dim=d0)) # 6
+        self.output_conditions.append(DepthTransformer(ch, 4, d0 // 2, context_dim=d0)) # 7
+        self.output_conditions.append(DepthTransformer(ch, 4, d0 // 2, context_dim=d0)) # 8
+
+    def forward(self, x, timesteps=None, context=None, source_dict=None, **kwargs):
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        h = x.type(self.dtype)
+        for index, module in enumerate(self.input_blocks):
+            h = module(h, emb, context)
+            hs.append(h)
+
+        h = self.middle_block(h, emb, context)
+        h = self.middle_conditions(h, context=source_dict[h.shape[-1]])
+
+        for index, module in enumerate(self.output_blocks):
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+            if index in self.output_b2c:
+                layer = self.output_conditions[self.output_b2c[index]]
+                h = layer(h, context=source_dict[h.shape[-1]])
+
+        h = h.type(x.dtype)
+        return self.out(h)
+
+    def get_trainable_parameters(self):
+        paras = [para for para in self.middle_conditions.parameters()] + [para for para in self.output_conditions.parameters()]
+        return paras

modelscope/models/cv/image_to_3d/ldm/models/diffusion/sync_dreamer_network.py

@@ -0,0 +1,186 @@
+import torch
+import torch.nn as nn
+
+class Image2DResBlockWithTV(nn.Module):
+    def __init__(self, dim, tdim, vdim):
+        super().__init__()
+        norm = lambda c: nn.GroupNorm(8, c)
+        self.time_embed = nn.Conv2d(tdim, dim, 1, 1)
+        self.view_embed = nn.Conv2d(vdim, dim, 1, 1)
+        self.conv = nn.Sequential(
+            norm(dim),
+            nn.SiLU(True),
+            nn.Conv2d(dim, dim, 3, 1, 1),
+            norm(dim),
+            nn.SiLU(True),
+            nn.Conv2d(dim, dim, 3, 1, 1),
+        )
+
+    def forward(self, x, t, v):
+        return x+self.conv(x+self.time_embed(t)+self.view_embed(v))
+
+
+class NoisyTargetViewEncoder(nn.Module):
+    def __init__(self, time_embed_dim, viewpoint_dim, run_dim=16, output_dim=8):
+        super().__init__()
+
+        self.init_conv = nn.Conv2d(4, run_dim, 3, 1, 1)
+        self.out_conv0 = Image2DResBlockWithTV(run_dim, time_embed_dim, viewpoint_dim)
+        self.out_conv1 = Image2DResBlockWithTV(run_dim, time_embed_dim, viewpoint_dim)
+        self.out_conv2 = Image2DResBlockWithTV(run_dim, time_embed_dim, viewpoint_dim)
+        self.final_out = nn.Sequential(
+            nn.GroupNorm(8, run_dim),
+            nn.SiLU(True),
+            nn.Conv2d(run_dim, output_dim, 3, 1, 1)
+        )
+
+    def forward(self, x, t, v):
+        B, DT = t.shape
+        t = t.view(B, DT, 1, 1)
+        B, DV = v.shape
+        v = v.view(B, DV, 1, 1)
+
+        x = self.init_conv(x)
+        x = self.out_conv0(x, t, v)
+        x = self.out_conv1(x, t, v)
+        x = self.out_conv2(x, t, v)
+        x = self.final_out(x)
+        return x
+
+class SpatialUpTimeBlock(nn.Module):
+    def __init__(self, x_in_dim, t_in_dim, out_dim):
+        super().__init__()
+        norm_act = lambda c: nn.GroupNorm(8, c)
+        self.t_conv = nn.Conv3d(t_in_dim, x_in_dim, 1, 1)  # 16
+        self.norm = norm_act(x_in_dim)
+        self.silu = nn.SiLU(True)
+        self.conv = nn.ConvTranspose3d(x_in_dim, out_dim, kernel_size=3, padding=1, output_padding=1, stride=2)
+
+    def forward(self, x, t):
+        x = x + self.t_conv(t)
+        return self.conv(self.silu(self.norm(x)))
+
+class SpatialTimeBlock(nn.Module):
+    def __init__(self, x_in_dim, t_in_dim, out_dim, stride):
+        super().__init__()
+        norm_act = lambda c: nn.GroupNorm(8, c)
+        self.t_conv = nn.Conv3d(t_in_dim, x_in_dim, 1, 1)  # 16
+        self.bn = norm_act(x_in_dim)
+        self.silu = nn.SiLU(True)
+        self.conv = nn.Conv3d(x_in_dim, out_dim, 3, stride=stride, padding=1)
+
+    def forward(self, x, t):
+        x = x + self.t_conv(t)
+        return self.conv(self.silu(self.bn(x)))
+
+class SpatialTime3DNet(nn.Module):
+        def __init__(self, time_dim=256, input_dim=128, dims=(32, 64, 128, 256)):
+            super().__init__()
+            d0, d1, d2, d3 = dims
+            dt = time_dim
+
+            self.init_conv = nn.Conv3d(input_dim, d0, 3, 1, 1)  # 32
+            self.conv0 = SpatialTimeBlock(d0, dt, d0, stride=1)
+
+            self.conv1 = SpatialTimeBlock(d0, dt, d1, stride=2)
+            self.conv2_0 = SpatialTimeBlock(d1, dt, d1, stride=1)
+            self.conv2_1 = SpatialTimeBlock(d1, dt, d1, stride=1)
+
+            self.conv3 = SpatialTimeBlock(d1, dt, d2, stride=2)
+            self.conv4_0 = SpatialTimeBlock(d2, dt, d2, stride=1)
+            self.conv4_1 = SpatialTimeBlock(d2, dt, d2, stride=1)
+
+            self.conv5 = SpatialTimeBlock(d2, dt, d3, stride=2)
+            self.conv6_0 = SpatialTimeBlock(d3, dt, d3, stride=1)
+            self.conv6_1 = SpatialTimeBlock(d3, dt, d3, stride=1)
+
+            self.conv7 = SpatialUpTimeBlock(d3, dt, d2)
+            self.conv8 = SpatialUpTimeBlock(d2, dt, d1)
+            self.conv9 = SpatialUpTimeBlock(d1, dt, d0)
+
+        def forward(self, x, t):
+            B, C = t.shape
+            t = t.view(B, C, 1, 1, 1)
+
+            x = self.init_conv(x)
+            conv0 = self.conv0(x, t)
+
+            x = self.conv1(conv0, t)
+            x = self.conv2_0(x, t)
+            conv2 = self.conv2_1(x, t)
+
+            x = self.conv3(conv2, t)
+            x = self.conv4_0(x, t)
+            conv4 = self.conv4_1(x, t)
+
+            x = self.conv5(conv4, t)
+            x = self.conv6_0(x, t)
+            x = self.conv6_1(x, t)
+
+            x = conv4 + self.conv7(x, t)
+            x = conv2 + self.conv8(x, t)
+            x = conv0 + self.conv9(x, t)
+            return x
+
+class FrustumTVBlock(nn.Module):
+    def __init__(self, x_dim, t_dim, v_dim, out_dim, stride):
+        super().__init__()
+        norm_act = lambda c: nn.GroupNorm(8, c)
+        self.t_conv = nn.Conv3d(t_dim, x_dim, 1, 1) # 16
+        self.v_conv = nn.Conv3d(v_dim, x_dim, 1, 1) # 16
+        self.bn = norm_act(x_dim)
+        self.silu = nn.SiLU(True)
+        self.conv = nn.Conv3d(x_dim, out_dim, 3, stride=stride, padding=1)
+
+    def forward(self, x, t, v):
+        x = x + self.t_conv(t) + self.v_conv(v)
+        return self.conv(self.silu(self.bn(x)))
+
+class FrustumTVUpBlock(nn.Module):
+    def __init__(self, x_dim, t_dim, v_dim, out_dim):
+        super().__init__()
+        norm_act = lambda c: nn.GroupNorm(8, c)
+        self.t_conv = nn.Conv3d(t_dim, x_dim, 1, 1) # 16
+        self.v_conv = nn.Conv3d(v_dim, x_dim, 1, 1) # 16
+        self.norm = norm_act(x_dim)
+        self.silu = nn.SiLU(True)
+        self.conv = nn.ConvTranspose3d(x_dim, out_dim, kernel_size=3, padding=1, output_padding=1, stride=2)
+
+    def forward(self, x, t, v):
+        x = x + self.t_conv(t) + self.v_conv(v)
+        return self.conv(self.silu(self.norm(x)))
+
+class FrustumTV3DNet(nn.Module):
+    def __init__(self, in_dim, t_dim, v_dim, dims=(32, 64, 128, 256)):
+        super().__init__()
+        self.conv0 = nn.Conv3d(in_dim, dims[0], 3, 1, 1) # 32
+
+        self.conv1 = FrustumTVBlock(dims[0], t_dim, v_dim, dims[1], 2)
+        self.conv2 = FrustumTVBlock(dims[1], t_dim, v_dim, dims[1], 1)
+
+        self.conv3 = FrustumTVBlock(dims[1], t_dim, v_dim, dims[2], 2)
+        self.conv4 = FrustumTVBlock(dims[2], t_dim, v_dim, dims[2], 1)
+
+        self.conv5 = FrustumTVBlock(dims[2], t_dim, v_dim, dims[3], 2)
+        self.conv6 = FrustumTVBlock(dims[3], t_dim, v_dim, dims[3], 1)
+
+        self.up0 = FrustumTVUpBlock(dims[3], t_dim, v_dim, dims[2])
+        self.up1 = FrustumTVUpBlock(dims[2], t_dim, v_dim, dims[1])
+        self.up2 = FrustumTVUpBlock(dims[1], t_dim, v_dim, dims[0])
+
+    def forward(self, x, t, v):
+        B,DT = t.shape
+        t = t.view(B,DT,1,1,1)
+        B,DV = v.shape
+        v = v.view(B,DV,1,1,1)
+
+        b, _, d, h, w = x.shape
+        x0 = self.conv0(x)
+        x1 = self.conv2(self.conv1(x0, t, v), t, v)
+        x2 = self.conv4(self.conv3(x1, t, v), t, v)
+        x3 = self.conv6(self.conv5(x2, t, v), t, v)
+
+        x2 = self.up0(x3, t, v) + x2
+        x1 = self.up1(x2, t, v) + x1
+        x0 = self.up2(x1, t, v) + x0
+        return {w: x0, w//2: x1, w//4: x2, w//8: x3}

modelscope/models/cv/image_to_3d/ldm/models/diffusion/sync_dreamer_utils.py

@@ -0,0 +1,103 @@
+import torch
+from kornia import create_meshgrid
+
+
+def project_and_normalize(ref_grid, src_proj, length):
+    """
+
+    @param ref_grid: b 3 n
+    @param src_proj: b 4 4
+    @param length:   int
+    @return:  b, n, 2
+    """
+    src_grid = src_proj[:, :3, :3] @ ref_grid + src_proj[:, :3, 3:] # b 3 n
+    div_val = src_grid[:, -1:]
+    div_val[div_val<1e-4] = 1e-4
+    src_grid = src_grid[:, :2] / div_val # divide by depth (b, 2, n)
+    src_grid[:, 0] = src_grid[:, 0]/((length - 1) / 2) - 1 # scale to -1~1
+    src_grid[:, 1] = src_grid[:, 1]/((length - 1) / 2) - 1 # scale to -1~1
+    src_grid = src_grid.permute(0, 2, 1) # (b, n, 2)
+    return src_grid
+
+
+def construct_project_matrix(x_ratio, y_ratio, Ks, poses):
+    """
+    @param x_ratio: float
+    @param y_ratio: float
+    @param Ks:      b,3,3
+    @param poses:   b,3,4
+    @return:
+    """
+    rfn = Ks.shape[0]
+    scale_m = torch.tensor([x_ratio, y_ratio, 1.0], dtype=torch.float32, device=Ks.device)
+    scale_m = torch.diag(scale_m)
+    ref_prj = scale_m[None, :, :] @ Ks @ poses  # rfn,3,4
+    pad_vals = torch.zeros([rfn, 1, 4], dtype=torch.float32, device=ref_prj.device)
+    pad_vals[:, :, 3] = 1.0
+    ref_prj = torch.cat([ref_prj, pad_vals], 1)  # rfn,4,4
+    return ref_prj
+
+def get_warp_coordinates(volume_xyz, warp_size, input_size, Ks, warp_pose):
+    B, _, D, H, W = volume_xyz.shape
+    ratio = warp_size / input_size
+    warp_proj = construct_project_matrix(ratio, ratio, Ks, warp_pose) # B,4,4
+    warp_coords = project_and_normalize(volume_xyz.view(B,3,D*H*W), warp_proj, warp_size).view(B, D, H, W, 2)
+    return warp_coords
+
+
+def create_target_volume(depth_size, volume_size, input_image_size, pose_target, K, near=None, far=None):
+    device, dtype = pose_target.device, pose_target.dtype
+
+    # compute a depth range on the unit sphere
+    H, W, D, B = volume_size, volume_size, depth_size, pose_target.shape[0]
+    if near is not None and far is not None :
+        # near, far b,1,h,w
+        depth_values = torch.linspace(0, 1, steps=depth_size).to(near.device).to(near.dtype) # d
+        depth_values = depth_values.view(1, D, 1, 1) # 1,d,1,1
+        depth_values = depth_values * (far - near) + near # b d h w
+        depth_values = depth_values.view(B, 1, D, H * W)
+    else:
+        near, far = near_far_from_unit_sphere_using_camera_poses(pose_target) # b 1
+        depth_values = torch.linspace(0, 1, steps=depth_size).to(near.device).to(near.dtype) # d
+        depth_values = depth_values[None,:,None] * (far[:,None,:] - near[:,None,:]) + near[:,None,:] # b d 1
+        depth_values = depth_values.view(B, 1, D, 1).expand(B, 1, D, H*W)
+
+    ratio = volume_size / input_image_size
+
+    # creat a grid on the target (reference) view
+    # H, W, D, B = volume_size, volume_size, depth_values.shape[1], depth_values.shape[0]
+
+    # creat mesh grid: note reference also means target
+    ref_grid = create_meshgrid(H, W, normalized_coordinates=False)  # (1, H, W, 2)
+    ref_grid = ref_grid.to(device).to(dtype)
+    ref_grid = ref_grid.permute(0, 3, 1, 2) # (1, 2, H, W)
+    ref_grid = ref_grid.reshape(1, 2, H*W) # (1, 2, H*W)
+    ref_grid = ref_grid.expand(B, -1, -1) # (B, 2, H*W)
+    ref_grid = torch.cat((ref_grid, torch.ones(B, 1, H*W, dtype=ref_grid.dtype, device=ref_grid.device)), dim=1) # (B, 3, H*W)
+    ref_grid = ref_grid.unsqueeze(2) * depth_values  # (B, 3, D, H*W)
+
+    # unproject to space and transfer to world coordinates.
+    Ks = K
+    ref_proj = construct_project_matrix(ratio, ratio, Ks, pose_target) # B,4,4
+    ref_proj_inv = torch.inverse(ref_proj) # B,4,4
+    ref_grid = ref_proj_inv[:,:3,:3] @ ref_grid.view(B,3,D*H*W) + ref_proj_inv[:,:3,3:] # B,3,3 @ B,3,DHW + B,3,1 => B,3,DHW
+    return ref_grid.reshape(B,3,D,H,W), depth_values.view(B,1,D,H,W)
+
+def near_far_from_unit_sphere_using_camera_poses(camera_poses):
+    """
+    @param camera_poses: b 3 4
+    @return:
+        near: b,1
+        far: b,1
+    """
+    R_w2c = camera_poses[..., :3, :3] # b 3 3
+    t_w2c = camera_poses[..., :3, 3:] # b 3 1
+    camera_origin = -R_w2c.permute(0,2,1) @ t_w2c   # b 3 1
+    # R_w2c.T @ (0,0,1) = z_dir
+    camera_orient = R_w2c.permute(0,2,1)[...,:3,2:3] # b 3 1
+    camera_origin, camera_orient = camera_origin[...,0], camera_orient[..., 0] # b 3
+    a = torch.sum(camera_orient ** 2, dim=-1, keepdim=True) # b 1
+    b = -torch.sum(camera_orient * camera_origin, dim=-1, keepdim=True) # b 1
+    mid = b / a # b 1
+    near, far = mid - 1.0, mid + 1.0
+    return near, far

modelscope/models/cv/image_to_3d/ldm/modules/attention.py

@@ -0,0 +1,336 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.util import checkpoint
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+# feedforward
+class ConvGEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Conv2d(dim_in, dim_out * 2, 1, 1, 0)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+        if exists(mask):
+            mask = mask>0
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        attn = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', attn, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(out)
+
+class BasicSpatialTransformer(nn.Module):
+    def __init__(self, dim, n_heads, d_head, context_dim=None, checkpoint=True):
+        super().__init__()
+        inner_dim = n_heads * d_head
+        self.proj_in = nn.Sequential(
+            nn.GroupNorm(8, dim),
+            nn.Conv2d(dim, inner_dim, kernel_size=1, stride=1, padding=0),
+            nn.GroupNorm(8, inner_dim),
+            nn.ReLU(True),
+        )
+        self.attn = CrossAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, context_dim=context_dim)  # is a self-attention if not self.disable_self_attn
+        self.out_conv = nn.Sequential(
+            nn.GroupNorm(8, inner_dim),
+            nn.ReLU(True),
+            nn.Conv2d(inner_dim, inner_dim, 1, 1),
+        )
+        self.proj_out = nn.Sequential(
+            nn.GroupNorm(8, inner_dim),
+            nn.ReLU(True),
+            zero_module(nn.Conv2d(inner_dim, dim, kernel_size=1, stride=1, padding=0)),
+        )
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context):
+        # input
+        b,_,h,w = x.shape
+        x_in = x
+        x = self.proj_in(x)
+
+        # attention
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        context = rearrange(context, 'b c h w -> b (h w) c').contiguous()
+        x = self.attn(x, context) + x
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+
+        # output
+        x = self.out_conv(x) + x
+        x = self.proj_out(x) + x_in
+        return x
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, disable_self_attn=False):
+        super().__init__()
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                                    context_dim=context_dim if self.disable_self_attn else None)  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
+                                    heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+class ConvFeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Conv2d(dim, inner_dim, 1, 1, 0),
+            nn.GELU()
+        ) if not glu else ConvGEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Conv2d(inner_dim, dim_out, 1, 1, 0)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(in_channels,
+                                 inner_dim,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim,
+                                   disable_self_attn=disable_self_attn)
+                for d in range(depth)]
+        )
+
+        self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                              in_channels,
+                                              kernel_size=1,
+                                              stride=1,
+                                              padding=0))
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        for block in self.transformer_blocks:
+            x = block(x, context=context)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        x = self.proj_out(x)
+        return x + x_in

modelscope/models/cv/image_to_3d/ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,835 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+
+from modelscope.models.cv.image_to_3d.ldm.util import instantiate_from_config
+from modelscope.models.cv.image_to_3d.ldm.modules.attention import LinearAttention
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z
+

modelscope/models/cv/image_to_3d/ldm/modules/diffusionmodules/openaimodel.py

@@ -0,0 +1,996 @@
+from abc import abstractmethod
+from functools import partial
+import math
+from typing import Iterable
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+)
+from modelscope.models.cv.image_to_3d.ldm.modules.attention import SpatialTransformer
+from modelscope.models.cv.image_to_3d.ldm.util import exists
+
+
+# dummy replace
+def convert_module_to_f16(x):
+    pass
+
+def convert_module_to_f32(x):
+    pass
+
+
+## go
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1)  # NC(HW)
+        x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+class TransposedUpsample(nn.Module):
+    'Learned 2x upsampling without padding'
+    def __init__(self, channels, out_channels=None, ks=5):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
+
+    def forward(self,x):
+        return self.up(x)
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm: # False
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = th.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return checkpoint(self._forward, (x,), self.parameters(), True)   # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
+        #return pt_checkpoint(self._forward, x)  # pytorch
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial ** 2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+        #self.num_res_blocks = num_res_blocks
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+            print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                  f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                  f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                  f"attention will still not be set.")  # todo: convert to warning
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        ) # 0
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions: # always True
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads_upsample,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) # N
+        emb = self.time_embed(t_emb) #
+
+        if self.num_classes is not None:
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context) # conv
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)
+
+
+class EncoderUNetModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    For usage, see UNet.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        pool="adaptive",
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=num_head_channels,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=num_head_channels,
+                use_new_attention_order=use_new_attention_order,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+        self.pool = pool
+        if pool == "adaptive":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                nn.AdaptiveAvgPool2d((1, 1)),
+                zero_module(conv_nd(dims, ch, out_channels, 1)),
+                nn.Flatten(),
+            )
+        elif pool == "attention":
+            assert num_head_channels != -1
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                AttentionPool2d(
+                    (image_size // ds), ch, num_head_channels, out_channels
+                ),
+            )
+        elif pool == "spatial":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                nn.ReLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        elif pool == "spatial_v2":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                normalization(2048),
+                nn.SiLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {pool} pooling")
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x K] Tensor of outputs.
+        """
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        results = []
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            if self.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))
+        h = self.middle_block(h, emb)
+        if self.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = th.cat(results, axis=-1)
+            return self.out(h)
+        else:
+            h = h.type(x.dtype)
+            return self.out(h)
+

modelscope/models/cv/image_to_3d/ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,267 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+from modelscope.models.cv.image_to_3d.ldm.util import instantiate_from_config
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

modelscope/models/cv/image_to_3d/ldm/modules/distributions/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

modelscope/models/cv/image_to_3d/ldm/modules/encoders/clip/__init__.py

@@ -0,0 +1 @@
+from .clip import *

modelscope/models/cv/image_to_3d/ldm/modules/encoders/clip/clip.py

@@ -0,0 +1,200 @@
+import hashlib
+import os
+import urllib
+import warnings
+from typing import Any, Union, List
+from pkg_resources import packaging
+
+import torch
+from PIL import Image
+from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
+from tqdm import tqdm
+
+from modelscope.models.cv.image_to_3d.ldm.modules.encoders.clip.model import build_model
+
+try:
+    from torchvision.transforms import InterpolationMode
+    BICUBIC = InterpolationMode.BICUBIC
+except ImportError:
+    BICUBIC = Image.BICUBIC
+
+
+if packaging.version.parse(torch.__version__) < packaging.version.parse("1.7.1"):
+    warnings.warn("PyTorch version 1.7.1 or higher is recommended")
+
+
+__all__ = ["available_models", "load"]
+
+_MODELS = {
+    "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
+    "RN50x16": "https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
+    "RN50x64": "https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
+    "ViT-B/32": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    "ViT-B/16": "https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
+    "ViT-L/14": "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
+    "ViT-L/14@336px": "https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt",
+}
+
+
+def _download(url: str, root: str):
+    os.makedirs(root, exist_ok=True)
+    filename = os.path.basename(url)
+
+    expected_sha256 = url.split("/")[-2]
+    download_target = os.path.join(root, filename)
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        if hashlib.sha256(open(download_target, "rb").read()).hexdigest() == expected_sha256:
+            return download_target
+        else:
+            warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file")
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True, unit_divisor=1024) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    if hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256:
+        raise RuntimeError("Model has been downloaded but the SHA256 checksum does not not match")
+
+    return download_target
+
+
+def _convert_image_to_rgb(image):
+    return image.convert("RGB")
+
+
+def _transform(n_px):
+    return Compose([
+        Resize(n_px, interpolation=BICUBIC),
+        CenterCrop(n_px),
+        _convert_image_to_rgb,
+        ToTensor(),
+        Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+    ])
+
+
+def available_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list(_MODELS.keys())
+
+
+def load(name: str, device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu", jit: bool = False, download_root: str = None):
+    """Load a CLIP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+
+    device : Union[str, torch.device]
+        The device to put the loaded model
+
+    jit : bool
+        Whether to load the optimized JIT model or more hackable non-JIT model (default).
+
+    download_root: str
+        path to download the model files; by default, it uses "~/.cache/clip"
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLIP model
+
+    preprocess : Callable[[PIL.Image], torch.Tensor]
+        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
+    """
+    if name in _MODELS:
+        model_path = _download(_MODELS[name], download_root or os.path.expanduser("~/.cache/clip"))
+    elif os.path.isfile(name):
+        model_path = name
+    else:
+        raise RuntimeError(f"Model {name} not found; available models = {available_models()}")
+
+    with open(model_path, 'rb') as opened_file:
+        try:
+            # loading JIT archive
+            model = torch.jit.load(opened_file, map_location=device if jit else "cpu").eval()
+            state_dict = None
+        except RuntimeError:
+            # loading saved state dict
+            if jit:
+                warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead")
+                jit = False
+            state_dict = torch.load(opened_file, map_location="cpu")
+
+    if not jit:
+        model = build_model(state_dict or model.state_dict()).to(device)
+        if str(device) == "cpu":
+            model.float()
+        return model, _transform(model.visual.input_resolution)
+
+    # patch the device names
+    device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[])
+    device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1]
+
+    def _node_get(node: torch._C.Node, key: str):
+        """Gets attributes of a node which is polymorphic over return type.
+        
+        From https://github.com/pytorch/pytorch/pull/82628
+        """
+        sel = node.kindOf(key)
+        return getattr(node, sel)(key)
+
+    def patch_device(module):
+        try:
+            graphs = [module.graph] if hasattr(module, "graph") else []
+        except RuntimeError:
+            graphs = []
+
+        if hasattr(module, "forward1"):
+            graphs.append(module.forward1.graph)
+
+        for graph in graphs:
+            for node in graph.findAllNodes("prim::Constant"):
+                if "value" in node.attributeNames() and str(_node_get(node, "value")).startswith("cuda"):
+                    node.copyAttributes(device_node)
+
+    model.apply(patch_device)
+    patch_device(model.encode_image)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 on CPU
+    if str(device) == "cpu":
+        float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[])
+        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
+        float_node = float_input.node()
+
+        def patch_float(module):
+            try:
+                graphs = [module.graph] if hasattr(module, "graph") else []
+            except RuntimeError:
+                graphs = []
+
+            if hasattr(module, "forward1"):
+                graphs.append(module.forward1.graph)
+
+            for graph in graphs:
+                for node in graph.findAllNodes("aten::to"):
+                    inputs = list(node.inputs())
+                    for i in [1, 2]:  # dtype can be the second or third argument to aten::to()
+                        if _node_get(inputs[i].node(), "value") == 5:
+                            inputs[i].node().copyAttributes(float_node)
+
+        model.apply(patch_float)
+        patch_float(model.encode_image)
+        patch_float(model.encode_text)
+
+        model.float()
+
+    return model, _transform(model.input_resolution.item())

modelscope/models/cv/image_to_3d/ldm/modules/encoders/clip/model.py

@@ -0,0 +1,436 @@
+from collections import OrderedDict
+from typing import Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu2 = nn.ReLU(inplace=True)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu3 = nn.ReLU(inplace=True)
+
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(OrderedDict([
+                ("-1", nn.AvgPool2d(stride)),
+                ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
+                ("1", nn.BatchNorm2d(planes * self.expansion))
+            ]))
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu1(self.bn1(self.conv1(x)))
+        out = self.relu2(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu3(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.flatten(start_dim=2).permute(2, 0, 1)  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1], key=x, value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False
+        )
+        return x.squeeze(0)
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, input_resolution=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.input_resolution = input_resolution
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        def stem(x):
+            x = self.relu1(self.bn1(self.conv1(x)))
+            x = self.relu2(self.bn2(self.conv2(x)))
+            x = self.relu3(self.bn3(self.conv3(x)))
+            x = self.avgpool(x)
+            return x
+
+        x = x.type(self.conv1.weight.dtype)
+        x = stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+class QuickGELU(nn.Module):
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_model * 4)),
+            ("gelu", QuickGELU()),
+            ("c_proj", nn.Linear(d_model * 4, d_model))
+        ]))
+        self.ln_2 = LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x: torch.Tensor):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
+
+    def forward(self, x: torch.Tensor):
+        return self.resblocks(x)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
+        self.ln_pre = LayerNorm(width)
+
+        self.transformer = Transformer(width, layers, heads)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+class CLIP(nn.Module):
+    def __init__(self,
+                 embed_dim: int,
+                 # vision
+                 image_resolution: int,
+                 vision_layers: Union[Tuple[int, int, int, int], int],
+                 vision_width: int,
+                 vision_patch_size: int,
+                 # text
+                 context_length: int,
+                 vocab_size: int,
+                 transformer_width: int,
+                 transformer_heads: int,
+                 transformer_layers: int
+                 ):
+        super().__init__()
+
+        self.context_length = context_length
+
+        if isinstance(vision_layers, (tuple, list)):
+            vision_heads = vision_width * 32 // 64
+            self.visual = ModifiedResNet(
+                layers=vision_layers,
+                output_dim=embed_dim,
+                heads=vision_heads,
+                input_resolution=image_resolution,
+                width=vision_width
+            )
+        else:
+            vision_heads = vision_width // 64
+            self.visual = VisionTransformer(
+                input_resolution=image_resolution,
+                patch_size=vision_patch_size,
+                width=vision_width,
+                layers=vision_layers,
+                heads=vision_heads,
+                output_dim=embed_dim
+            )
+
+        self.transformer = Transformer(
+            width=transformer_width,
+            layers=transformer_layers,
+            heads=transformer_heads,
+            attn_mask=self.build_attention_mask()
+        )
+
+        self.vocab_size = vocab_size
+        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
+        self.ln_final = LayerNorm(transformer_width)
+
+        self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+        self.initialize_parameters()
+
+    def initialize_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+
+        if isinstance(self.visual, ModifiedResNet):
+            if self.visual.attnpool is not None:
+                std = self.visual.attnpool.c_proj.in_features ** -0.5
+                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)
+
+            for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
+                for name, param in resnet_block.named_parameters():
+                    if name.endswith("bn3.weight"):
+                        nn.init.zeros_(param)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    @property
+    def dtype(self):
+        return self.visual.conv1.weight.dtype
+
+    def encode_image(self, image):
+        return self.visual(image.type(self.dtype))
+
+    def encode_text(self, text):
+        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.type(self.dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x).type(self.dtype)
+
+        # x.shape = [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+
+        return x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image)
+        text_features = self.encode_text(text)
+
+        # normalized features
+        image_features = image_features / image_features.norm(dim=1, keepdim=True)
+        text_features = text_features / text_features.norm(dim=1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_image = logit_scale * image_features @ text_features.t()
+        logits_per_text = logits_per_image.t()
+
+        # shape = [global_batch_size, global_batch_size]
+        return logits_per_image, logits_per_text
+
+
+def convert_weights(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+def build_model(state_dict: dict):
+    vit = "visual.proj" in state_dict
+
+    if vit:
+        vision_width = state_dict["visual.conv1.weight"].shape[0]
+        vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+        image_resolution = vision_patch_size * grid_size
+    else:
+        counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
+        vision_layers = tuple(counts)
+        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
+        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
+        vision_patch_size = None
+        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
+        image_resolution = output_width * 32
+
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks")))
+
+    model = CLIP(
+        embed_dim,
+        image_resolution, vision_layers, vision_width, vision_patch_size,
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
+    )
+
+    for key in ["input_resolution", "context_length", "vocab_size"]:
+        if key in state_dict:
+            del state_dict[key]
+
+    convert_weights(model)
+    model.load_state_dict(state_dict)
+    return model.eval()

modelscope/models/cv/image_to_3d/ldm/modules/encoders/clip/simple_tokenizer.py

@@ -0,0 +1,132 @@
+import gzip
+import html
+import os
+from functools import lru_cache
+
+import ftfy
+import regex as re
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe()):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
+        merges = merges[1:49152-256-2+1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v+'</w>' for v in vocab]
+        for merge in merges:
+            vocab.append(''.join(merge))
+        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + ( token[-1] + '</w>',)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token+'</w>'
+
+        while True:
+            bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word)-1 and word[i+1] == second:
+                    new_word.append(first+second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = ' '.join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
+            bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = ''.join([self.decoder[token] for token in tokens])
+        text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
+        return text

modelscope/models/cv/image_to_3d/ldm/modules/encoders/modules.py

@@ -0,0 +1,551 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from functools import partial
+import kornia
+
+from modelscope.models.cv.image_to_3d.ldm.modules.x_transformer import Encoder, TransformerWrapper  # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test
+from modelscope.models.cv.image_to_3d.ldm.util import default
+# import clip
+from modelscope.models.cv.image_to_3d.ldm.modules.encoders import clip
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+class IdentityEncoder(AbstractEncoder):
+
+    def encode(self, x):
+        return x
+
+class FaceClipEncoder(AbstractEncoder):
+    def __init__(self, augment=True, retreival_key=None):
+        super().__init__()
+        self.encoder = FrozenCLIPImageEmbedder()
+        self.augment = augment
+        self.retreival_key = retreival_key
+
+    def forward(self, img):
+        encodings = []
+        with torch.no_grad():
+            x_offset = 125
+            if self.retreival_key:
+                # Assumes retrieved image are packed into the second half of channels
+                face = img[:,3:,190:440,x_offset:(512-x_offset)]
+                other = img[:,:3,...].clone()
+            else:
+                face = img[:,:,190:440,x_offset:(512-x_offset)]
+                other = img.clone()
+
+            if self.augment:
+                face = K.RandomHorizontalFlip()(face)
+
+            other[:,:,190:440,x_offset:(512-x_offset)] *= 0
+            encodings = [
+                self.encoder.encode(face),
+                self.encoder.encode(other),
+            ]
+
+        return torch.cat(encodings, dim=1)
+
+    def encode(self, img):
+        if isinstance(img, list):
+            # Uncondition
+            return torch.zeros((1, 2, 768), device=self.encoder.model.visual.conv1.weight.device)
+
+        return self(img)
+
+class FaceIdClipEncoder(AbstractEncoder):
+    def __init__(self):
+        super().__init__()
+        self.encoder = FrozenCLIPImageEmbedder()
+        for p in self.encoder.parameters():
+            p.requires_grad = False
+        self.id = FrozenFaceEncoder("/home/jpinkney/code/stable-diffusion/model_ir_se50.pth", augment=True)
+
+    def forward(self, img):
+        encodings = []
+        with torch.no_grad():
+            face = kornia.geometry.resize(img, (256, 256),
+                            interpolation='bilinear', align_corners=True)
+
+            other = img.clone()
+            other[:,:,184:452,122:396] *= 0
+            encodings = [
+                self.id.encode(face),
+                self.encoder.encode(other),
+            ]
+
+        return torch.cat(encodings, dim=1)
+
+    def encode(self, img):
+        if isinstance(img, list):
+            # Uncondition
+            return torch.zeros((1, 2, 768), device=self.encoder.model.visual.conv1.weight.device)
+
+        return self(img)
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class'):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        c = self.embedding(c)
+        return c
+
+
+class TransformerEmbedder(AbstractEncoder):
+    """Some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"):
+        super().__init__()
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer))
+
+    def forward(self, tokens):
+        tokens = tokens.to(self.device)  # meh
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, x):
+        return self(x)
+
+
+class BERTTokenizer(AbstractEncoder):
+    """ Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)"""
+    def __init__(self, device="cuda", vq_interface=True, max_length=77):
+        super().__init__()
+        from transformers import BertTokenizerFast  # TODO: add to reuquirements
+        self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
+        self.device = device
+        self.vq_interface = vq_interface
+        self.max_length = max_length
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        return tokens
+
+    @torch.no_grad()
+    def encode(self, text):
+        tokens = self(text)
+        if not self.vq_interface:
+            return tokens
+        return None, None, [None, None, tokens]
+
+    def decode(self, text):
+        return text
+
+
+class BERTEmbedder(AbstractEncoder):
+    """Uses the BERT tokenizr model and add some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77,
+                 device="cuda",use_tokenizer=True, embedding_dropout=0.0):
+        super().__init__()
+        self.use_tknz_fn = use_tokenizer
+        if self.use_tknz_fn:
+            self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len)
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer),
+                                              emb_dropout=embedding_dropout)
+
+    def forward(self, text):
+        if self.use_tknz_fn:
+            tokens = self.tknz_fn(text)#.to(self.device)
+        else:
+            tokens = text
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, text):
+        # output of length 77
+        return self(text)
+
+
+from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class FrozenT5Embedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version, cache_dir='/apdcephfs/private_rondyliu/projects/huggingface_models')
+        self.transformer = T5EncoderModel.from_pretrained(version, cache_dir='/apdcephfs/private_rondyliu/projects/huggingface_models')
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+from modelscope.models.cv.image_to_3d.ldm.thirdp.psp.id_loss import IDFeatures
+import kornia.augmentation as K
+
+class FrozenFaceEncoder(AbstractEncoder):
+    def __init__(self, model_path, augment=False):
+        super().__init__()
+        self.loss_fn = IDFeatures(model_path)
+        # face encoder is frozen
+        for p in self.loss_fn.parameters():
+            p.requires_grad = False
+        # Mapper is trainable
+        self.mapper = torch.nn.Linear(512, 768)
+        p = 0.25
+        if augment:
+            self.augment = K.AugmentationSequential(
+                K.RandomHorizontalFlip(p=0.5),
+                K.RandomEqualize(p=p),
+                # K.RandomPlanckianJitter(p=p),
+                # K.RandomPlasmaBrightness(p=p),
+                # K.RandomPlasmaContrast(p=p),
+                # K.ColorJiggle(0.02, 0.2, 0.2, p=p),
+            )
+        else:
+            self.augment = False
+
+    def forward(self, img):
+        if isinstance(img, list):
+            # Uncondition
+            return torch.zeros((1, 1, 768), device=self.mapper.weight.device)
+
+        if self.augment is not None:
+            # Transforms require 0-1
+            img = self.augment((img + 1)/2)
+            img = 2*img - 1
+
+        feat = self.loss_fn(img, crop=True)
+        feat = self.mapper(feat.unsqueeze(1))
+        return feat
+
+    def encode(self, img):
+        return self(img)
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from huggingface)"""
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(version, cache_dir='/apdcephfs/private_rondyliu/projects/huggingface_models')
+        self.transformer = CLIPTextModel.from_pretrained(version, cache_dir='/apdcephfs/private_rondyliu/projects/huggingface_models')
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+import torch.nn.functional as F
+from transformers import CLIPVisionModel
+class ClipImageProjector(AbstractEncoder):
+    """
+        Uses the CLIP image encoder.
+        """
+    def __init__(self, version="openai/clip-vit-large-patch14", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        self.model = CLIPVisionModel.from_pretrained(version)
+        self.model.train()
+        self.max_length = max_length   # TODO: typical value?
+        self.antialias = True
+        self.mapper = torch.nn.Linear(1024, 768)
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        null_cond = self.get_null_cond(version, max_length)
+        self.register_buffer('null_cond', null_cond)
+
+    @torch.no_grad()
+    def get_null_cond(self, version, max_length):
+        device = self.mean.device
+        embedder = FrozenCLIPEmbedder(version=version, device=device, max_length=max_length)
+        null_cond = embedder([""])
+        return null_cond
+
+    def preprocess(self, x):
+        # Expects inputs in the range -1, 1
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic',align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        if isinstance(x, list):
+            return self.null_cond
+        # x is assumed to be in range [-1,1]
+        x = self.preprocess(x)
+        outputs = self.model(pixel_values=x)
+        last_hidden_state = outputs.last_hidden_state
+        last_hidden_state = self.mapper(last_hidden_state)
+        return F.pad(last_hidden_state, [0,0, 0,self.max_length-last_hidden_state.shape[1], 0,0])
+
+    def encode(self, im):
+        return self(im)
+
+class ProjectedFrozenCLIPEmbedder(AbstractEncoder):
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        self.embedder = FrozenCLIPEmbedder(version=version, device=device, max_length=max_length)
+        self.projection = torch.nn.Linear(768, 768)
+
+    def forward(self, text):
+        z = self.embedder(text)
+        return self.projection(z)
+
+    def encode(self, text):
+        return self(text)
+
+class FrozenCLIPImageEmbedder(AbstractEncoder):
+    """
+        Uses the CLIP image encoder.
+        Not actually frozen... If you want that set cond_stage_trainable=False in cfg
+        """
+    def __init__(
+            self,
+            model='ViT-L/14',
+            jit=False,
+            device='cpu',
+            antialias=False,
+        ):
+        super().__init__()
+        self.model, _ = clip.load(name=model, device=device, jit=jit)
+        # We don't use the text part so delete it
+        del self.model.transformer
+        self.antialias = antialias
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+
+    def preprocess(self, x):
+        # Expects inputs in the range -1, 1
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic',align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        # x is assumed to be in range [-1,1]
+        if isinstance(x, list):
+            # [""] denotes condition dropout for ucg
+            device = self.model.visual.conv1.weight.device
+            return torch.zeros(1, 768, device=device)
+        return self.model.encode_image(self.preprocess(x)).float()
+
+    def encode(self, im):
+        return self(im).unsqueeze(1)
+
+from torchvision import transforms
+import random
+
+class FrozenCLIPImageMutliEmbedder(AbstractEncoder):
+    """
+        Uses the CLIP image encoder.
+        Not actually frozen... If you want that set cond_stage_trainable=False in cfg
+        """
+    def __init__(
+            self,
+            model='ViT-L/14',
+            jit=False,
+            device='cpu',
+            antialias=True,
+            max_crops=5,
+        ):
+        super().__init__()
+        self.model, _ = clip.load(name=model, device=device, jit=jit)
+        # We don't use the text part so delete it
+        del self.model.transformer
+        self.antialias = antialias
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        self.max_crops = max_crops
+
+    def preprocess(self, x):
+
+        # Expects inputs in the range -1, 1
+        randcrop = transforms.RandomResizedCrop(224, scale=(0.085, 1.0), ratio=(1,1))
+        max_crops = self.max_crops
+        patches = []
+        crops = [randcrop(x) for _ in range(max_crops)]
+        patches.extend(crops)
+        x = torch.cat(patches, dim=0)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        # x is assumed to be in range [-1,1]
+        if isinstance(x, list):
+            # [""] denotes condition dropout for ucg
+            device = self.model.visual.conv1.weight.device
+            return torch.zeros(1, self.max_crops, 768, device=device)
+        batch_tokens = []
+        for im in x:
+            patches = self.preprocess(im.unsqueeze(0))
+            tokens = self.model.encode_image(patches).float()
+            for t in tokens:
+                if random.random() < 0.1:
+                    t *= 0
+            batch_tokens.append(tokens.unsqueeze(0))
+
+        return torch.cat(batch_tokens, dim=0)
+
+    def encode(self, im):
+        return self(im)
+
+class SpatialRescaler(nn.Module):
+    def __init__(self,
+                 n_stages=1,
+                 method='bilinear',
+                 multiplier=0.5,
+                 in_channels=3,
+                 out_channels=None,
+                 bias=False):
+        super().__init__()
+        self.n_stages = n_stages
+        assert self.n_stages >= 0
+        assert method in ['nearest','linear','bilinear','trilinear','bicubic','area']
+        self.multiplier = multiplier
+        self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
+        self.remap_output = out_channels is not None
+        if self.remap_output:
+            print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.')
+            self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias)
+
+    def forward(self,x):
+        for stage in range(self.n_stages):
+            x = self.interpolator(x, scale_factor=self.multiplier)
+
+
+        if self.remap_output:
+            x = self.channel_mapper(x)
+        return x
+
+    def encode(self, x):
+        return self(x)
+
+
+from modelscope.models.cv.image_to_3d.ldm.util import instantiate_from_config
+from modelscope.models.cv.image_to_3d.ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+
+
+class LowScaleEncoder(nn.Module):
+    def __init__(self, model_config, linear_start, linear_end, timesteps=1000, max_noise_level=250, output_size=64,
+                 scale_factor=1.0):
+        super().__init__()
+        self.max_noise_level = max_noise_level
+        self.model = instantiate_from_config(model_config)
+        self.augmentation_schedule = self.register_schedule(timesteps=timesteps, linear_start=linear_start,
+                                                            linear_end=linear_end)
+        self.out_size = output_size
+        self.scale_factor = scale_factor
+
+    def register_schedule(self, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                   cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def forward(self, x):
+        z = self.model.encode(x).sample()
+        z = z * self.scale_factor
+        noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        z = self.q_sample(z, noise_level)
+        if self.out_size is not None:
+            z = torch.nn.functional.interpolate(z, size=self.out_size, mode="nearest")  # TODO: experiment with mode
+        # z = z.repeat_interleave(2, -2).repeat_interleave(2, -1)
+        return z, noise_level
+
+    def decode(self, z):
+        z = z / self.scale_factor
+        return self.model.decode(z)
+
+
+if __name__ == "__main__":
+    from ldm.util import count_params
+    sentences = ["a hedgehog drinking a whiskey", "der mond ist aufgegangen", "Ein Satz mit vielen Sonderzeichen: äöü ß ?! : 'xx-y/@s'"]
+    model = FrozenT5Embedder(version="google/t5-v1_1-xl").cuda()
+    count_params(model, True)
+    z = model(sentences)
+    print(z.shape)
+
+    model = FrozenCLIPEmbedder().cuda()
+    count_params(model, True)
+    z = model(sentences)
+    print(z.shape)
+
+    print("done.")

modelscope/models/cv/image_to_3d/ldm/modules/x_transformer.py

@@ -0,0 +1,641 @@
+"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+from functools import partial
+from inspect import isfunction
+from collections import namedtuple
+from einops import rearrange, repeat, reduce
+
+# constants
+
+DEFAULT_DIM_HEAD = 64
+
+Intermediates = namedtuple('Intermediates', [
+    'pre_softmax_attn',
+    'post_softmax_attn'
+])
+
+LayerIntermediates = namedtuple('Intermediates', [
+    'hiddens',
+    'attn_intermediates'
+])
+
+
+class AbsolutePositionalEmbedding(nn.Module):
+    def __init__(self, dim, max_seq_len):
+        super().__init__()
+        self.emb = nn.Embedding(max_seq_len, dim)
+        self.init_()
+
+    def init_(self):
+        nn.init.normal_(self.emb.weight, std=0.02)
+
+    def forward(self, x):
+        n = torch.arange(x.shape[1], device=x.device)
+        return self.emb(n)[None, :, :]
+
+
+class FixedPositionalEmbedding(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer('inv_freq', inv_freq)
+
+    def forward(self, x, seq_dim=1, offset=0):
+        t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
+        sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
+        emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
+        return emb[None, :, :]
+
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def always(val):
+    def inner(*args, **kwargs):
+        return val
+    return inner
+
+
+def not_equals(val):
+    def inner(x):
+        return x != val
+    return inner
+
+
+def equals(val):
+    def inner(x):
+        return x == val
+    return inner
+
+
+def max_neg_value(tensor):
+    return -torch.finfo(tensor.dtype).max
+
+
+# keyword argument helpers
+
+def pick_and_pop(keys, d):
+    values = list(map(lambda key: d.pop(key), keys))
+    return dict(zip(keys, values))
+
+
+def group_dict_by_key(cond, d):
+    return_val = [dict(), dict()]
+    for key in d.keys():
+        match = bool(cond(key))
+        ind = int(not match)
+        return_val[ind][key] = d[key]
+    return (*return_val,)
+
+
+def string_begins_with(prefix, str):
+    return str.startswith(prefix)
+
+
+def group_by_key_prefix(prefix, d):
+    return group_dict_by_key(partial(string_begins_with, prefix), d)
+
+
+def groupby_prefix_and_trim(prefix, d):
+    kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
+    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
+    return kwargs_without_prefix, kwargs
+
+
+# classes
+class Scale(nn.Module):
+    def __init__(self, value, fn):
+        super().__init__()
+        self.value = value
+        self.fn = fn
+
+    def forward(self, x, **kwargs):
+        x, *rest = self.fn(x, **kwargs)
+        return (x * self.value, *rest)
+
+
+class Rezero(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+        self.g = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x, **kwargs):
+        x, *rest = self.fn(x, **kwargs)
+        return (x * self.g, *rest)
+
+
+class ScaleNorm(nn.Module):
+    def __init__(self, dim, eps=1e-5):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(1))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
+        return x / norm.clamp(min=self.eps) * self.g
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-8):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
+        return x / norm.clamp(min=self.eps) * self.g
+
+
+class Residual(nn.Module):
+    def forward(self, x, residual):
+        return x + residual
+
+
+class GRUGating(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.gru = nn.GRUCell(dim, dim)
+
+    def forward(self, x, residual):
+        gated_output = self.gru(
+            rearrange(x, 'b n d -> (b n) d'),
+            rearrange(residual, 'b n d -> (b n) d')
+        )
+
+        return gated_output.reshape_as(x)
+
+
+# feedforward
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+# attention.
+class Attention(nn.Module):
+    def __init__(
+            self,
+            dim,
+            dim_head=DEFAULT_DIM_HEAD,
+            heads=8,
+            causal=False,
+            mask=None,
+            talking_heads=False,
+            sparse_topk=None,
+            use_entmax15=False,
+            num_mem_kv=0,
+            dropout=0.,
+            on_attn=False
+    ):
+        super().__init__()
+        if use_entmax15:
+            raise NotImplementedError("Check out entmax activation instead of softmax activation!")
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        self.causal = causal
+        self.mask = mask
+
+        inner_dim = dim_head * heads
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(dim, inner_dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+        # talking heads
+        self.talking_heads = talking_heads
+        if talking_heads:
+            self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
+            self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
+
+        # explicit topk sparse attention
+        self.sparse_topk = sparse_topk
+
+        # entmax
+        #self.attn_fn = entmax15 if use_entmax15 else F.softmax
+        self.attn_fn = F.softmax
+
+        # add memory key / values
+        self.num_mem_kv = num_mem_kv
+        if num_mem_kv > 0:
+            self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
+            self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
+
+        # attention on attention
+        self.attn_on_attn = on_attn
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)
+
+    def forward(
+            self,
+            x,
+            context=None,
+            mask=None,
+            context_mask=None,
+            rel_pos=None,
+            sinusoidal_emb=None,
+            prev_attn=None,
+            mem=None
+    ):
+        b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
+        kv_input = default(context, x)
+
+        q_input = x
+        k_input = kv_input
+        v_input = kv_input
+
+        if exists(mem):
+            k_input = torch.cat((mem, k_input), dim=-2)
+            v_input = torch.cat((mem, v_input), dim=-2)
+
+        if exists(sinusoidal_emb):
+            # in shortformer, the query would start at a position offset depending on the past cached memory
+            offset = k_input.shape[-2] - q_input.shape[-2]
+            q_input = q_input + sinusoidal_emb(q_input, offset=offset)
+            k_input = k_input + sinusoidal_emb(k_input)
+
+        q = self.to_q(q_input)
+        k = self.to_k(k_input)
+        v = self.to_v(v_input)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
+
+        input_mask = None
+        if any(map(exists, (mask, context_mask))):
+            q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
+            k_mask = q_mask if not exists(context) else context_mask
+            k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
+            q_mask = rearrange(q_mask, 'b i -> b () i ()')
+            k_mask = rearrange(k_mask, 'b j -> b () () j')
+            input_mask = q_mask * k_mask
+
+        if self.num_mem_kv > 0:
+            mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
+            k = torch.cat((mem_k, k), dim=-2)
+            v = torch.cat((mem_v, v), dim=-2)
+            if exists(input_mask):
+                input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
+        mask_value = max_neg_value(dots)
+
+        if exists(prev_attn):
+            dots = dots + prev_attn
+
+        pre_softmax_attn = dots
+
+        if talking_heads:
+            dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
+
+        if exists(rel_pos):
+            dots = rel_pos(dots)
+
+        if exists(input_mask):
+            dots.masked_fill_(~input_mask, mask_value)
+            del input_mask
+
+        if self.causal:
+            i, j = dots.shape[-2:]
+            r = torch.arange(i, device=device)
+            mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
+            mask = F.pad(mask, (j - i, 0), value=False)
+            dots.masked_fill_(mask, mask_value)
+            del mask
+
+        if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
+            top, _ = dots.topk(self.sparse_topk, dim=-1)
+            vk = top[..., -1].unsqueeze(-1).expand_as(dots)
+            mask = dots < vk
+            dots.masked_fill_(mask, mask_value)
+            del mask
+
+        attn = self.attn_fn(dots, dim=-1)
+        post_softmax_attn = attn
+
+        attn = self.dropout(attn)
+
+        if talking_heads:
+            attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
+
+        out = einsum('b h i j, b h j d -> b h i d', attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+
+        intermediates = Intermediates(
+            pre_softmax_attn=pre_softmax_attn,
+            post_softmax_attn=post_softmax_attn
+        )
+
+        return self.to_out(out), intermediates
+
+
+class AttentionLayers(nn.Module):
+    def __init__(
+            self,
+            dim,
+            depth,
+            heads=8,
+            causal=False,
+            cross_attend=False,
+            only_cross=False,
+            use_scalenorm=False,
+            use_rmsnorm=False,
+            use_rezero=False,
+            rel_pos_num_buckets=32,
+            rel_pos_max_distance=128,
+            position_infused_attn=False,
+            custom_layers=None,
+            sandwich_coef=None,
+            par_ratio=None,
+            residual_attn=False,
+            cross_residual_attn=False,
+            macaron=False,
+            pre_norm=True,
+            gate_residual=False,
+            **kwargs
+    ):
+        super().__init__()
+        ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
+        attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
+
+        dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
+
+        self.dim = dim
+        self.depth = depth
+        self.layers = nn.ModuleList([])
+
+        self.has_pos_emb = position_infused_attn
+        self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
+        self.rotary_pos_emb = always(None)
+
+        assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
+        self.rel_pos = None
+
+        self.pre_norm = pre_norm
+
+        self.residual_attn = residual_attn
+        self.cross_residual_attn = cross_residual_attn
+
+        norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
+        norm_class = RMSNorm if use_rmsnorm else norm_class
+        norm_fn = partial(norm_class, dim)
+
+        norm_fn = nn.Identity if use_rezero else norm_fn
+        branch_fn = Rezero if use_rezero else None
+
+        if cross_attend and not only_cross:
+            default_block = ('a', 'c', 'f')
+        elif cross_attend and only_cross:
+            default_block = ('c', 'f')
+        else:
+            default_block = ('a', 'f')
+
+        if macaron:
+            default_block = ('f',) + default_block
+
+        if exists(custom_layers):
+            layer_types = custom_layers
+        elif exists(par_ratio):
+            par_depth = depth * len(default_block)
+            assert 1 < par_ratio <= par_depth, 'par ratio out of range'
+            default_block = tuple(filter(not_equals('f'), default_block))
+            par_attn = par_depth // par_ratio
+            depth_cut = par_depth * 2 // 3  # 2 / 3 attention layer cutoff suggested by PAR paper
+            par_width = (depth_cut + depth_cut // par_attn) // par_attn
+            assert len(default_block) <= par_width, 'default block is too large for par_ratio'
+            par_block = default_block + ('f',) * (par_width - len(default_block))
+            par_head = par_block * par_attn
+            layer_types = par_head + ('f',) * (par_depth - len(par_head))
+        elif exists(sandwich_coef):
+            assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
+            layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
+        else:
+            layer_types = default_block * depth
+
+        self.layer_types = layer_types
+        self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
+
+        for layer_type in self.layer_types:
+            if layer_type == 'a':
+                layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
+            elif layer_type == 'c':
+                layer = Attention(dim, heads=heads, **attn_kwargs)
+            elif layer_type == 'f':
+                layer = FeedForward(dim, **ff_kwargs)
+                layer = layer if not macaron else Scale(0.5, layer)
+            else:
+                raise Exception(f'invalid layer type {layer_type}')
+
+            if isinstance(layer, Attention) and exists(branch_fn):
+                layer = branch_fn(layer)
+
+            if gate_residual:
+                residual_fn = GRUGating(dim)
+            else:
+                residual_fn = Residual()
+
+            self.layers.append(nn.ModuleList([
+                norm_fn(),
+                layer,
+                residual_fn
+            ]))
+
+    def forward(
+            self,
+            x,
+            context=None,
+            mask=None,
+            context_mask=None,
+            mems=None,
+            return_hiddens=False
+    ):
+        hiddens = []
+        intermediates = []
+        prev_attn = None
+        prev_cross_attn = None
+
+        mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
+
+        for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
+            is_last = ind == (len(self.layers) - 1)
+
+            if layer_type == 'a':
+                hiddens.append(x)
+                layer_mem = mems.pop(0)
+
+            residual = x
+
+            if self.pre_norm:
+                x = norm(x)
+
+            if layer_type == 'a':
+                out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
+                                   prev_attn=prev_attn, mem=layer_mem)
+            elif layer_type == 'c':
+                out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
+            elif layer_type == 'f':
+                out = block(x)
+
+            x = residual_fn(out, residual)
+
+            if layer_type in ('a', 'c'):
+                intermediates.append(inter)
+
+            if layer_type == 'a' and self.residual_attn:
+                prev_attn = inter.pre_softmax_attn
+            elif layer_type == 'c' and self.cross_residual_attn:
+                prev_cross_attn = inter.pre_softmax_attn
+
+            if not self.pre_norm and not is_last:
+                x = norm(x)
+
+        if return_hiddens:
+            intermediates = LayerIntermediates(
+                hiddens=hiddens,
+                attn_intermediates=intermediates
+            )
+
+            return x, intermediates
+
+        return x
+
+
+class Encoder(AttentionLayers):
+    def __init__(self, **kwargs):
+        assert 'causal' not in kwargs, 'cannot set causality on encoder'
+        super().__init__(causal=False, **kwargs)
+
+
+
+class TransformerWrapper(nn.Module):
+    def __init__(
+            self,
+            *,
+            num_tokens,
+            max_seq_len,
+            attn_layers,
+            emb_dim=None,
+            max_mem_len=0.,
+            emb_dropout=0.,
+            num_memory_tokens=None,
+            tie_embedding=False,
+            use_pos_emb=True
+    ):
+        super().__init__()
+        assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
+
+        dim = attn_layers.dim
+        emb_dim = default(emb_dim, dim)
+
+        self.max_seq_len = max_seq_len
+        self.max_mem_len = max_mem_len
+        self.num_tokens = num_tokens
+
+        self.token_emb = nn.Embedding(num_tokens, emb_dim)
+        self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
+                    use_pos_emb and not attn_layers.has_pos_emb) else always(0)
+        self.emb_dropout = nn.Dropout(emb_dropout)
+
+        self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
+        self.attn_layers = attn_layers
+        self.norm = nn.LayerNorm(dim)
+
+        self.init_()
+
+        self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
+
+        # memory tokens (like [cls]) from Memory Transformers paper
+        num_memory_tokens = default(num_memory_tokens, 0)
+        self.num_memory_tokens = num_memory_tokens
+        if num_memory_tokens > 0:
+            self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
+
+            # let funnel encoder know number of memory tokens, if specified
+            if hasattr(attn_layers, 'num_memory_tokens'):
+                attn_layers.num_memory_tokens = num_memory_tokens
+
+    def init_(self):
+        nn.init.normal_(self.token_emb.weight, std=0.02)
+
+    def forward(
+            self,
+            x,
+            return_embeddings=False,
+            mask=None,
+            return_mems=False,
+            return_attn=False,
+            mems=None,
+            **kwargs
+    ):
+        b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
+        x = self.token_emb(x)
+        x += self.pos_emb(x)
+        x = self.emb_dropout(x)
+
+        x = self.project_emb(x)
+
+        if num_mem > 0:
+            mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
+            x = torch.cat((mem, x), dim=1)
+
+            # auto-handle masking after appending memory tokens
+            if exists(mask):
+                mask = F.pad(mask, (num_mem, 0), value=True)
+
+        x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
+        x = self.norm(x)
+
+        mem, x = x[:, :num_mem], x[:, num_mem:]
+
+        out = self.to_logits(x) if not return_embeddings else x
+
+        if return_mems:
+            hiddens = intermediates.hiddens
+            new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
+            new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
+            return out, new_mems
+
+        if return_attn:
+            attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
+            return out, attn_maps
+
+        return out
+

modelscope/models/cv/image_to_3d/ldm/thirdp/psp/helpers.py

@@ -0,0 +1,121 @@
+# https://github.com/eladrich/pixel2style2pixel
+
+from collections import namedtuple
+import torch
+from torch.nn import Conv2d, BatchNorm2d, PReLU, ReLU, Sigmoid, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module
+
+"""
+ArcFace implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
+"""
+
+
+class Flatten(Module):
+	def forward(self, input):
+		return input.view(input.size(0), -1)
+
+
+def l2_norm(input, axis=1):
+	norm = torch.norm(input, 2, axis, True)
+	output = torch.div(input, norm)
+	return output
+
+
+class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])):
+	""" A named tuple describing a ResNet block. """
+
+
+def get_block(in_channel, depth, num_units, stride=2):
+	return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)]
+
+
+def get_blocks(num_layers):
+	if num_layers == 50:
+		blocks = [
+			get_block(in_channel=64, depth=64, num_units=3),
+			get_block(in_channel=64, depth=128, num_units=4),
+			get_block(in_channel=128, depth=256, num_units=14),
+			get_block(in_channel=256, depth=512, num_units=3)
+		]
+	elif num_layers == 100:
+		blocks = [
+			get_block(in_channel=64, depth=64, num_units=3),
+			get_block(in_channel=64, depth=128, num_units=13),
+			get_block(in_channel=128, depth=256, num_units=30),
+			get_block(in_channel=256, depth=512, num_units=3)
+		]
+	elif num_layers == 152:
+		blocks = [
+			get_block(in_channel=64, depth=64, num_units=3),
+			get_block(in_channel=64, depth=128, num_units=8),
+			get_block(in_channel=128, depth=256, num_units=36),
+			get_block(in_channel=256, depth=512, num_units=3)
+		]
+	else:
+		raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers))
+	return blocks
+
+
+class SEModule(Module):
+	def __init__(self, channels, reduction):
+		super(SEModule, self).__init__()
+		self.avg_pool = AdaptiveAvgPool2d(1)
+		self.fc1 = Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False)
+		self.relu = ReLU(inplace=True)
+		self.fc2 = Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False)
+		self.sigmoid = Sigmoid()
+
+	def forward(self, x):
+		module_input = x
+		x = self.avg_pool(x)
+		x = self.fc1(x)
+		x = self.relu(x)
+		x = self.fc2(x)
+		x = self.sigmoid(x)
+		return module_input * x
+
+
+class bottleneck_IR(Module):
+	def __init__(self, in_channel, depth, stride):
+		super(bottleneck_IR, self).__init__()
+		if in_channel == depth:
+			self.shortcut_layer = MaxPool2d(1, stride)
+		else:
+			self.shortcut_layer = Sequential(
+				Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+				BatchNorm2d(depth)
+			)
+		self.res_layer = Sequential(
+			BatchNorm2d(in_channel),
+			Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth),
+			Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth)
+		)
+
+	def forward(self, x):
+		shortcut = self.shortcut_layer(x)
+		res = self.res_layer(x)
+		return res + shortcut
+
+
+class bottleneck_IR_SE(Module):
+	def __init__(self, in_channel, depth, stride):
+		super(bottleneck_IR_SE, self).__init__()
+		if in_channel == depth:
+			self.shortcut_layer = MaxPool2d(1, stride)
+		else:
+			self.shortcut_layer = Sequential(
+				Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+				BatchNorm2d(depth)
+			)
+		self.res_layer = Sequential(
+			BatchNorm2d(in_channel),
+			Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False),
+			PReLU(depth),
+			Conv2d(depth, depth, (3, 3), stride, 1, bias=False),
+			BatchNorm2d(depth),
+			SEModule(depth, 16)
+		)
+
+	def forward(self, x):
+		shortcut = self.shortcut_layer(x)
+		res = self.res_layer(x)
+		return res + shortcut

modelscope/models/cv/image_to_3d/ldm/thirdp/psp/id_loss.py

@@ -0,0 +1,23 @@
+# https://github.com/eladrich/pixel2style2pixel
+import torch
+from torch import nn
+from modelscope.models.cv.image_to_3d.ldm.thirdp.psp.model_irse import Backbone
+
+
+class IDFeatures(nn.Module):
+    def __init__(self, model_path):
+        super(IDFeatures, self).__init__()
+        print('Loading ResNet ArcFace')
+        self.facenet = Backbone(input_size=112, num_layers=50, drop_ratio=0.6, mode='ir_se')
+        self.facenet.load_state_dict(torch.load(model_path, map_location="cpu"))
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
+        self.facenet.eval()
+
+    def forward(self, x, crop=False):
+        # Not sure of the image range here
+        if crop:
+            x = torch.nn.functional.interpolate(x, (256, 256), mode="area")
+            x = x[:, :, 35:223, 32:220]
+        x = self.face_pool(x)
+        x_feats = self.facenet(x)
+        return x_feats

modelscope/models/cv/image_to_3d/ldm/thirdp/psp/model_irse.py

@@ -0,0 +1,86 @@
+# https://github.com/eladrich/pixel2style2pixel
+
+from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Dropout, Sequential, Module
+from modelscope.models.cv.image_to_3d.ldm.thirdp.psp.helpers import get_blocks, Flatten, bottleneck_IR, bottleneck_IR_SE, l2_norm
+
+"""
+Modified Backbone implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
+"""
+
+
+class Backbone(Module):
+	def __init__(self, input_size, num_layers, mode='ir', drop_ratio=0.4, affine=True):
+		super(Backbone, self).__init__()
+		assert input_size in [112, 224], "input_size should be 112 or 224"
+		assert num_layers in [50, 100, 152], "num_layers should be 50, 100 or 152"
+		assert mode in ['ir', 'ir_se'], "mode should be ir or ir_se"
+		blocks = get_blocks(num_layers)
+		if mode == 'ir':
+			unit_module = bottleneck_IR
+		elif mode == 'ir_se':
+			unit_module = bottleneck_IR_SE
+		self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
+									  BatchNorm2d(64),
+									  PReLU(64))
+		if input_size == 112:
+			self.output_layer = Sequential(BatchNorm2d(512),
+			                               Dropout(drop_ratio),
+			                               Flatten(),
+			                               Linear(512 * 7 * 7, 512),
+			                               BatchNorm1d(512, affine=affine))
+		else:
+			self.output_layer = Sequential(BatchNorm2d(512),
+			                               Dropout(drop_ratio),
+			                               Flatten(),
+			                               Linear(512 * 14 * 14, 512),
+			                               BatchNorm1d(512, affine=affine))
+
+		modules = []
+		for block in blocks:
+			for bottleneck in block:
+				modules.append(unit_module(bottleneck.in_channel,
+										   bottleneck.depth,
+										   bottleneck.stride))
+		self.body = Sequential(*modules)
+
+	def forward(self, x):
+		x = self.input_layer(x)
+		x = self.body(x)
+		x = self.output_layer(x)
+		return l2_norm(x)
+
+
+def IR_50(input_size):
+	"""Constructs a ir-50 model."""
+	model = Backbone(input_size, num_layers=50, mode='ir', drop_ratio=0.4, affine=False)
+	return model
+
+
+def IR_101(input_size):
+	"""Constructs a ir-101 model."""
+	model = Backbone(input_size, num_layers=100, mode='ir', drop_ratio=0.4, affine=False)
+	return model
+
+
+def IR_152(input_size):
+	"""Constructs a ir-152 model."""
+	model = Backbone(input_size, num_layers=152, mode='ir', drop_ratio=0.4, affine=False)
+	return model
+
+
+def IR_SE_50(input_size):
+	"""Constructs a ir_se-50 model."""
+	model = Backbone(input_size, num_layers=50, mode='ir_se', drop_ratio=0.4, affine=False)
+	return model
+
+
+def IR_SE_101(input_size):
+	"""Constructs a ir_se-101 model."""
+	model = Backbone(input_size, num_layers=100, mode='ir_se', drop_ratio=0.4, affine=False)
+	return model
+
+
+def IR_SE_152(input_size):
+	"""Constructs a ir_se-152 model."""
+	model = Backbone(input_size, num_layers=152, mode='ir_se', drop_ratio=0.4, affine=False)
+	return model

modelscope/models/cv/image_to_3d/ldm/util.py

@@ -0,0 +1,276 @@
+import importlib
+
+import torchvision
+import torch
+from torch import optim
+import numpy as np
+
+from inspect import isfunction
+from PIL import Image, ImageDraw, ImageFont
+
+import os
+import numpy as np
+import matplotlib.pyplot as plt
+from PIL import Image
+import torch
+import time
+import cv2
+import PIL
+
+def pil_rectangle_crop(im):
+    width, height = im.size   # Get dimensions
+    
+    if width <= height:
+        left = 0
+        right = width
+        top = (height - width)/2
+        bottom = (height + width)/2
+    else:
+        
+        top = 0
+        bottom = height
+        left = (width - height) / 2
+        bottom = (width + height) / 2
+
+    # Crop the center of the image
+    im = im.crop((left, top, right, bottom))
+    return im
+
+def add_margin(pil_img, color=0, size=256):
+    width, height = pil_img.size
+    result = Image.new(pil_img.mode, (size, size), color)
+    result.paste(pil_img, ((size - width) // 2, (size - height) // 2))
+    return result
+
+
+def create_carvekit_interface():
+    from carvekit.api.high import HiInterface
+    # Check doc strings for more information
+    interface = HiInterface(object_type="object",  # Can be "object" or "hairs-like".
+                            batch_size_seg=5,
+                            batch_size_matting=1,
+                            device='cuda' if torch.cuda.is_available() else 'cpu',
+                            seg_mask_size=640,  # Use 640 for Tracer B7 and 320 for U2Net
+                            matting_mask_size=2048,
+                            trimap_prob_threshold=231,
+                            trimap_dilation=30,
+                            trimap_erosion_iters=5,
+                            fp16=False)
+
+    return interface
+
+
+def load_and_preprocess(interface, input_im):
+    '''
+    :param input_im (PIL Image).
+    :return image (H, W, 3) array in [0, 1].
+    '''
+    # See https://github.com/Ir1d/image-background-remove-tool
+    image = input_im.convert('RGB')
+
+    image_without_background = interface([image])[0]
+    image_without_background = np.array(image_without_background)
+    est_seg = image_without_background > 127
+    image = np.array(image)
+    foreground = est_seg[:, : , -1].astype(np.bool_)
+    image[~foreground] = [255., 255., 255.]
+    x, y, w, h = cv2.boundingRect(foreground.astype(np.uint8))
+    image = image[y:y+h, x:x+w, :]
+    image = PIL.Image.fromarray(np.array(image))
+    
+    # resize image such that long edge is 512
+    image.thumbnail([200, 200], Image.LANCZOS)
+    image = add_margin(image, (255, 255, 255), size=256)
+    image = np.array(image)
+    
+    return image
+
+
+def log_txt_as_img(wh, xc, size=10):
+    # wh a tuple of (width, height)
+    # xc a list of captions to plot
+    b = len(xc)
+    txts = list()
+    for bi in range(b):
+        txt = Image.new("RGB", wh, color="white")
+        draw = ImageDraw.Draw(txt)
+        font = ImageFont.truetype('data/DejaVuSans.ttf', size=size)
+        nc = int(40 * (wh[0] / 256))
+        lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc))
+
+        try:
+            draw.text((0, 0), lines, fill="black", font=font)
+        except UnicodeEncodeError:
+            print("Cant encode string for logging. Skipping.")
+
+        txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
+        txts.append(txt)
+    txts = np.stack(txts)
+    txts = torch.tensor(txts)
+    return txts
+
+
+def ismap(x):
+    if not isinstance(x, torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] > 3)
+
+
+def isimage(x):
+    if not isinstance(x,torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
+
+
+def exists(x):
+    return x is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def mean_flat(tensor):
+    """
+    https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
+    return total_params
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    print(module)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+class AdamWwithEMAandWings(optim.Optimizer):
+    # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298
+    def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8,  # TODO: check hyperparameters before using
+                 weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999,   # ema decay to match previous code
+                 ema_power=1., param_names=()):
+        """AdamW that saves EMA versions of the parameters."""
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0.0 <= weight_decay:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0.0 <= ema_decay <= 1.0:
+            raise ValueError("Invalid ema_decay value: {}".format(ema_decay))
+        defaults = dict(lr=lr, betas=betas, eps=eps,
+                        weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay,
+                        ema_power=ema_power, param_names=param_names)
+        super().__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super().__setstate__(state)
+        for group in self.param_groups:
+            group.setdefault('amsgrad', False)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+        Args:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            params_with_grad = []
+            grads = []
+            exp_avgs = []
+            exp_avg_sqs = []
+            ema_params_with_grad = []
+            state_sums = []
+            max_exp_avg_sqs = []
+            state_steps = []
+            amsgrad = group['amsgrad']
+            beta1, beta2 = group['betas']
+            ema_decay = group['ema_decay']
+            ema_power = group['ema_power']
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                params_with_grad.append(p)
+                if p.grad.is_sparse:
+                    raise RuntimeError('AdamW does not support sparse gradients')
+                grads.append(p.grad)
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state['step'] = 0
+                    # Exponential moving average of gradient values
+                    state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    if amsgrad:
+                        # Maintains max of all exp. moving avg. of sq. grad. values
+                        state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of parameter values
+                    state['param_exp_avg'] = p.detach().float().clone()
+
+                exp_avgs.append(state['exp_avg'])
+                exp_avg_sqs.append(state['exp_avg_sq'])
+                ema_params_with_grad.append(state['param_exp_avg'])
+
+                if amsgrad:
+                    max_exp_avg_sqs.append(state['max_exp_avg_sq'])
+
+                # update the steps for each param group update
+                state['step'] += 1
+                # record the step after step update
+                state_steps.append(state['step'])
+
+            optim._functional.adamw(params_with_grad,
+                    grads,
+                    exp_avgs,
+                    exp_avg_sqs,
+                    max_exp_avg_sqs,
+                    state_steps,
+                    amsgrad=amsgrad,
+                    beta1=beta1,
+                    beta2=beta2,
+                    lr=group['lr'],
+                    weight_decay=group['weight_decay'],
+                    eps=group['eps'],
+                    maximize=False)
+
+            cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power)
+            for param, ema_param in zip(params_with_grad, ema_params_with_grad):
+                ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay)
+
+        return loss

modelscope/outputs/outputs.py

@@ -69,6 +69,7 @@ class OutputKeys(object):
     PCD12 = 'pcd12'
     PCD12_ALIGN = 'pcd12_align'
     TBOUNDS = 'tbounds'
+    MV_IMGS = 'MViews'
 
 
 OutputTypes = {
@@ -132,6 +133,7 @@ OutputTypes = {
     OutputKeys.PCD12: np.ndarray,
     OutputKeys.PCD12_ALIGN: np.ndarray,
     OutputKeys.TBOUNDS: Dict,
+    OutputKeys.MV_IMGS: List[np.ndarray],
 }
 
 OutputTypeSchema = {
@@ -426,6 +428,15 @@ OutputTypeSchema = {
     OutputKeys.TBOUNDS: {
         'type': 'object'
     },
+    OutputKeys.MV_IMGS: {
+        'type': 'array',
+        'items': {
+            'type': 'array',
+            'items': {
+                'type': 'number'
+            }
+        }
+    },
 }
 
 TASK_OUTPUTS = {
@@ -1632,6 +1643,7 @@ TASK_OUTPUTS = {
     #    "output_imgs": np.ndarray list with shape [[height, width, 3], ...]
     # }
     Tasks.image_view_transform: [OutputKeys.OUTPUT_IMGS],
+    Tasks.image_to_3d: [OutputKeys.MV_IMGS]
 }
 
 

modelscope/pipelines/cv/__init__.py

@@ -41,6 +41,7 @@ if TYPE_CHECKING:
     from .image_super_resolution_pasd_pipeline import ImageSuperResolutionPASDPipeline
     from .image_to_image_generate_pipeline import Image2ImageGenerationPipeline
     from .image_to_image_translation_pipeline import Image2ImageTranslationPipeline
+
     from .image_inpainting_pipeline import ImageInpaintingPipeline
     from .image_paintbyexample_pipeline import ImagePaintbyexamplePipeline
     from .product_retrieval_embedding_pipeline import ProductRetrievalEmbeddingPipeline
@@ -107,6 +108,7 @@ if TYPE_CHECKING:
     from .image_human_parsing_pipeline import ImageHumanParsingPipeline
     from .nerf_recon_acc_pipeline import NeRFReconAccPipeline
     from .nerf_recon_4k_pipeline import NeRFRecon4KPipeline
+    from .image_to_3d_pipeline import Image23DPipeline
     from .surface_recon_common_pipeline import SurfaceReconCommonPipeline
     from .controllable_image_generation_pipeline import ControllableImageGenerationPipeline
     from .image_bts_depth_estimation_pipeline import ImageBTSDepthEstimationPipeline
@@ -163,6 +165,7 @@ else:
         ['ProductRetrievalEmbeddingPipeline'],
         'live_category_pipeline': ['LiveCategoryPipeline'],
         'image_to_image_generate_pipeline': ['Image2ImageGenerationPipeline'],
+        'image_to_3d_pipeline': ['Image23DPipeline'],
         'image_inpainting_pipeline': ['ImageInpaintingPipeline'],
         'image_paintbyexample_pipeline': ['ImagePaintbyexamplePipeline'],
         'ocr_detection_pipeline': ['OCRDetectionPipeline'],
@@ -269,6 +272,7 @@ else:
         'image_human_parsing_pipeline': ['ImageHumanParsingPipeline'],
         'nerf_recon_acc_pipeline': ['NeRFReconAccPipeline'],
         'nerf_recon_4k_pipeline': ['NeRFRecon4KPipeline'],
+        'nerf_recon_img_to_mv_pipeline': ['NeRFReconImgToMVPipeline'],
         'surface_recon_common_pipeline': ['SurfaceReconCommonPipeline'],
         'controllable_image_generation_pipeline': [
             'ControllableImageGenerationPipeline'

modelscope/pipelines/cv/image_to_3d_pipeline.py

@@ -0,0 +1,125 @@
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import os.path as osp
+from typing import Any, Dict
+import rembg
+import cv2
+import numpy as np
+import PIL
+import torch
+import torch.nn.functional as F
+import torchvision.transforms as T
+import torchvision.transforms.functional as TF
+from PIL import Image
+from torchvision.utils import save_image
+from omegaconf import OmegaConf
+# import modelscope.models.cv.image_to_image_generation.data as data
+# import modelscope.models.cv.image_to_image_generation.models as models
+# import modelscope.models.cv.image_to_image_generation.ops as ops
+from modelscope.metainfo import Pipelines
+# from modelscope.models.cv.image_to_3d.model import UNet
+# from modelscope.models.cv.image_to_image_generation.models.clip import \
+#     VisionTransformer
+
+from modelscope.models.cv.image_to_3d.ldm.models.diffusion.sync_dreamer import SyncMultiviewDiffusion
+from modelscope.models.cv.image_to_3d.ldm.util import instantiate_from_config, add_margin
+
+from modelscope.outputs import OutputKeys
+from modelscope.pipelines.base import Input, Pipeline
+from modelscope.pipelines.builder import PIPELINES
+from modelscope.preprocessors import LoadImage
+from modelscope.utils.config import Config
+from modelscope.utils.constant import ModelFile, Tasks
+from modelscope.utils.logger import get_logger
+
+logger = get_logger()
+
+# Load Syncdreamer Model
+def load_model(cfg, ckpt, strict=True):
+    config = OmegaConf.load(cfg)
+    model = instantiate_from_config(config.model)
+    print(f'loading model from {ckpt} ...')
+    ckpt = torch.load(ckpt,map_location='cpu')
+    model.load_state_dict(ckpt['state_dict'],strict=strict)
+    model = model.cuda().eval()
+    return model
+
+# Prepare Syncdreamer Input
+def prepare_inputs(image_input, elevation_input, crop_size=-1, image_size=256):
+    image_input[:,:,:3] = image_input[:,:,:3][:,:,::-1]
+    image_input = Image.fromarray(image_input)
+    if crop_size!=-1:
+        alpha_np = np.asarray(image_input)[:, :, 3]
+        coords = np.stack(np.nonzero(alpha_np), 1)[:, (1, 0)]
+        min_x, min_y = np.min(coords, 0)
+        max_x, max_y = np.max(coords, 0)
+        ref_img_ = image_input.crop((min_x, min_y, max_x, max_y))
+        h, w = ref_img_.height, ref_img_.width
+        scale = crop_size / max(h, w)
+        h_, w_ = int(scale * h), int(scale * w)
+        ref_img_ = ref_img_.resize((w_, h_), resample=Image.BICUBIC)
+        image_input = add_margin(ref_img_, size=image_size)
+    else:
+        image_input = add_margin(image_input, size=max(image_input.height, image_input.width))
+        image_input = image_input.resize((image_size, image_size), resample=Image.BICUBIC)
+
+    image_input = np.asarray(image_input)
+    image_input = image_input.astype(np.float32) / 255.0
+    ref_mask = image_input[:, :, 3:]
+    image_input[:, :, :3] = image_input[:, :, :3] * ref_mask + 1 - ref_mask  # white background
+    image_input = image_input[:, :, :3] * 2.0 - 1.0
+    image_input = torch.from_numpy(image_input.astype(np.float32))
+    elevation_input = torch.from_numpy(np.asarray([np.deg2rad(elevation_input)], np.float32))
+    return {"input_image": image_input, "input_elevation": elevation_input}
+
+@PIPELINES.register_module(
+    Tasks.image_to_3d,
+    module_name=Pipelines.image_to_3d)
+class Image23DPipeline(Pipeline):
+
+    def __init__(self, model: str, **kwargs):
+        """
+        use `model` to create a image-to-3d generation pipeline
+        Args:
+            model: model id on modelscope hub.
+        """
+        super().__init__(model=model)
+        config_path = osp.join(self.model, ModelFile.CONFIGURATION)
+        logger.info(f'loading config from {config_path}')
+        self.cfg = Config.from_file(config_path)
+        # print(config_path)
+        if torch.cuda.is_available():
+            self._device = torch.device('cuda')
+        else:
+            self._device = torch.device('cpu')
+        ckpt = config_path.replace("configuration.json", "syncdreamer-pretrain.ckpt")
+        self.model = load_model(config_path.replace("configuration.json", "syncdreamer.yaml"), ckpt).to(self._device)
+        # os.system("pip install -r {}".format(config_path.replace("configuration.json", "requirements.txt")))
+        # assert isinstance(self.model, SyncMultiviewDiffusion)
+
+    def preprocess(self, input: Input) -> Dict[str, Any]:
+        
+        result = rembg.remove(Image.open(input))
+        print(type(result))
+        img = np.array(result)
+        img[:,:,:3] = img[:,:,:3][:,:,::-1]
+        # img = cv2.imread(input)
+        data = prepare_inputs(img, elevation_input=10, crop_size=200, image_size=256)
+        
+        for k,v in data.items():
+            data[k] = v.unsqueeze(0).cuda()
+            data[k] = torch.repeat_interleave(data[k], 1, dim=0) # only one sample
+        return data
+
+    @torch.no_grad()
+    def forward(self, input: Dict[str, Any]) -> Dict[str, Any]:
+        x_sample = self.model.sample(input, 2.0, 8)
+
+        B, N, _, H, W = x_sample.shape
+        x_sample = (torch.clamp(x_sample,max=1.0,min=-1.0) + 1) * 0.5
+        x_sample = x_sample.permute(0,1,3,4,2).cpu().numpy() * 255
+        x_sample = x_sample.astype(np.uint8)
+        show_in_im2 = [Image.fromarray(x_sample[0,ni]) for ni in range(N)]
+        return {'MViews':show_in_im2}
+
+    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        return inputs

modelscope/utils/constant.py

@@ -169,6 +169,9 @@ class CVTasks(object):
     human3d_animation = 'human3d-animation'
     image_control_3d_portrait = 'image-control-3d-portrait'
 
+    # 3d generation
+    image_to_3d = 'image-to-3d'
+
     # vision efficient tuning
     vision_efficient_tuning = 'vision-efficient-tuning'
 

tests/pipelines/test_image_to_3d.py

@@ -0,0 +1,41 @@
+# Copyright 2021-2022 The Alibaba Fundamental Vision Team Authors. All rights reserved.
+import unittest
+
+import numpy as np
+from PIL import Image
+from modelscope.outputs import OutputKeys
+from modelscope.pipelines import pipeline
+from modelscope.pipelines.base import Pipeline
+from modelscope.utils.constant import Tasks
+from modelscope.utils.logger import get_logger
+from modelscope.utils.test_utils import test_level
+
+logger = get_logger()
+
+
+class ImageTo3DTest(unittest.TestCase):
+
+    def setUp(self) -> None:
+        self.model_id = 'Damo_XR_Lab/Syncdreamer'
+        self.input = {
+            'input_path': 'data/test/images/basketball.png',
+        }
+
+    def pipeline_inference(self, pipeline: Pipeline, input: str):
+        result = pipeline(input['input_path'])
+        np_content = []
+        for idx,img in enumerate(result['MViews']):
+            np_content.append(np.array(result['MViews'][idx]))
+
+        np_content = np.concatenate(np_content, axis=1)
+        Image.fromarray(np_content).save("./concat.png")
+
+    @unittest.skipUnless(test_level() >= 0, 'skip test in current test level')
+    def test_run_modelhub(self):
+        image_to_3d = pipeline(
+            Tasks.image_to_3d, model=self.model_id, revision='v1.0.1')
+        self.pipeline_inference(image_to_3d, self.input)
+
+
+if __name__ == '__main__':
+    unittest.main()