first release

FollowYourPose

@@ -1 +0,0 @@
-Subproject commit 40a333f7f1e3940903916c419b725a4a17f348a1

FollowYourPose/.gitignore

@@ -0,0 +1,51 @@
+
+/checkpoints
+
+/data
+
+/sim_matrix
+
+/figs
+
+/log
+
+
+*.log
+
+*.sh 
+
+*.pth
+
+*.jpg
+
+*.og
+
+/logs
+
+/stable_diffusion.egg-info
+
+/__pycache__
+
+/outputs
+
+*.pyc
+
+/models
+
+*.ckpt
+
+/src
+
+/newsd_weight
+
+*.yaml
+
+*.png
+
+*.yaml
+
+*.txt
+
+/data
+
+/others

FollowYourPose/followyourpose/data/hdvila.py

@@ -0,0 +1,245 @@
+# import os
+# import random
+# from abc import abstractmethod
+# import math
+# import pandas as pd
+# import av
+# import cv2
+# import decord
+# import numpy as np
+# import torch
+# from PIL import Image
+# from torch.utils.data import Dataset
+# from torchvision import transforms
+# from decord import VideoReader, cpu
+# import torchvision.transforms._transforms_video as transforms_video
+# from torchvision.transforms.functional import to_tensor
+# from collections import OrderedDict
+# import time
+# import csv
+
+# class HDVilaDataset(Dataset):
+#     """
+#     HDVila Dataset.
+#     Assumes webvid data is structured as follows.
+#     HDVila/
+#         part_N/ # 0-10
+#             video_clips/        ($page_dir) 
+#                 1.mp4           (videoid.mp4)
+#                 ...
+#                 5000.mp4
+#             ...
+#     """
+#     def __init__(self,
+#                 video_path,
+#                 width=512,
+#                 height=512,
+#                 n_sample_frames=8,
+#                 dataset_set="train",
+#                 prompt=None,
+#                 sample_frame_rate=2,
+#                 sample_start_idx=0,
+#                 accelerator=None,
+#                 ):
+        
+#         try:
+#             host_gpu_num = accelerator.num_processes
+#             host_num = 1
+#             all_rank = host_gpu_num * host_num
+#             global_rank = accelerator.local_process_index
+#         except:
+#             pass
+#         print('dataset rank:', global_rank, ' / ',all_rank, ' ')
+        
+#         self.data_dir = '/apdcephfs_cq3/share_1290939/0_public_datasets/hd-vila-100m'
+#         if dataset_set=='train':
+#             self.text_name = 'caption_rm2048_train.csv'
+#         else:
+#             self.text_name = 'caption_2048_val_new.csv'
+#         self.meta_path = os.path.join(self.data_dir, self.text_name)
+#         # text_name = 'caption_2048_val_new.csv'
+        
+#         spatial_transform = 'resize_center_crop'
+#         resolution=width
+#         load_raw_resolution=True
+#         # frame_stride=2
+#         video_length= n_sample_frames
+#         fps_max=None
+#         load_resize_keep_ratio=False
+        
+        
+        
+#         self.global_rank = global_rank
+#         self.all_rank = all_rank
+#         # self.subsample = subsample
+#         self.video_length = video_length
+#         self.resolution = [resolution, resolution] if isinstance(resolution, int) else resolution
+#         self.frame_stride = sample_frame_rate
+#         self.load_raw_resolution = load_raw_resolution
+#         self.fps_max = fps_max
+#         self.load_resize_keep_ratio = load_resize_keep_ratio
+#         print('start load meta data')
+#         self._load_metadata()
+#         print('load meta data done!!!')
+#         if spatial_transform is not None:
+#             if spatial_transform == "random_crop":
+#                 self.spatial_transform = transforms_video.RandomCropVideo(crop_resolution)
+#             elif spatial_transform == "resize_center_crop":
+#                 assert(self.resolution[0] == self.resolution[1])
+#                 self.spatial_transform = transforms.Compose([
+#                     transforms.Resize(resolution),
+#                     transforms_video.CenterCropVideo(resolution),
+#                     ])
+#             elif spatial_transform == "center_crop":
+#                 self.spatial_transform = transforms_video.CenterCropVideo(resolution)
+#             else:
+#                 raise NotImplementedError
+#         else:
+#             self.spatial_transform = None
+    
+#     def _load_metadata(self):
+#         # clip_id frame_id caption 
+#         last_clip_id = ''
+#         self.metadata = []
+#         start_time = time.time()
+#         caption_path = self.meta_path
+#         count=-1
+#         total_count = 8854264 #8856312 - 2048
+
+#         with open(caption_path, 'r',encoding="utf-8") as csvfile: #41s
+#             reader = csv.DictReader(csvfile)
+#             for row in reader:
+#                 if row['clip_id'] != last_clip_id:
+#                     count+=1
+#                     if count >= (total_count // self.all_rank)*self.all_rank: # drop last
+#                         break
+#                     last_clip_id = row['clip_id']
+#                     if count % self.all_rank == self.global_rank:
+#                         self.metadata.append([('%02d'%int(row['part_id']))+row['clip_id']]) 
+#                         self.metadata[-1].append([row['caption']])
+#                 else:
+#                     if count % self.all_rank == self.global_rank:
+#                         self.metadata[-1][-1].append(row['caption'])
+#             # caption_data = pd.read_csv(caption_path) # use time 26+264s
+
+#             # for index,row in caption_data.iterrows():
+#             #     if row['clip_id'] != last_clip_id:
+#             #         last_clip_id = row['clip_id']
+#             #         meta_data[('%02d'%part_id)+row['clip_id']] = [row['caption']]
+#             #     else:
+#             #         meta_data[('%02d'%part_id)+row['clip_id']].append(row['caption'])
+#         end_time = time.time()
+#         print('load %d - %d items use time: %.1f;' % (len(self.metadata), count, end_time-start_time))
+#         # self.metadata=meta_data
+
+    
+#     def _get_video_path(self, sample):
+#         part_id = int(sample[0][:2])
+#         clip_id = sample[0][2:]
+#         video_path = os.path.join(self.data_dir,'part_%d' % part_id, 'video_clips', clip_id)
+#         return video_path
+    
+#     def __getitem__(self, index):
+#         while True:
+            
+#             index = index % len(self.metadata)
+#             sample = self.metadata[index]
+#             video_path = self._get_video_path(sample)
+             
+#             try:
+#                 if self.load_raw_resolution:
+#                     video_reader = VideoReader(video_path, ctx=cpu(0))
+#                 elif self.load_resize_keep_ratio:
+#                     # resize scale is according to the short side
+#                     h, w, c = VideoReader(video_path, ctx=cpu(0))[0].shape
+#                     if h < w:
+#                         scale = h / self.resolution[0]
+#                     else:
+#                         scale = w / self.resolution[1]
+
+#                     h = math.ceil(h / scale)
+#                     w = math.ceil(w / scale)
+#                     video_reader = VideoReader(video_path, ctx=cpu(0), width=w, height=h)
+#                 else:
+#                     video_reader = VideoReader(video_path, ctx=cpu(0), width=self.resolution[1], height=self.resolution[0])
+#                 if len(video_reader) < self.video_length:
+#                     print(f"video length ({len(video_reader)}) is smaller than target length({self.video_length})")
+#                     index += 1
+#                     continue
+#                 else:
+#                     pass
+#             except:
+#                 index += 1
+#                 print(f"Load video failed! path = {video_path}")
+#                 continue
+#             fps_ori = video_reader.get_avg_fps()
+
+#             fs = self.frame_stride
+#             allf = len(video_reader)
+#             if self.frame_stride != 1:
+#                 all_frames = list(range(0, len(video_reader), self.frame_stride))
+#                 if len(all_frames) < self.video_length:
+#                     fs = len(video_reader) // self.video_length
+#                     assert(fs != 0)
+#                     all_frames = list(range(0, len(video_reader), fs))
+#             else:
+#                 all_frames = list(range(len(video_reader)))
+            
+#             # select a random clip
+#             rand_idx = random.randint(0, len(all_frames) - self.video_length)
+#             frame_indices = all_frames[rand_idx:rand_idx+self.video_length]
+#             try:
+#                 frames = video_reader.get_batch(frame_indices)
+#                 break
+#             except:
+#                 print(f"Get frames failed! path = {video_path}")
+#                 index += 1
+#                 continue
+
+#         assert(frames.shape[0] == self.video_length),f'{len(frames)}, self.video_length={self.video_length}'
+#         frames = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float() # [t,h,w,c] -> [c,t,h,w]
+       
+#         if self.spatial_transform is not None:
+#             frames = self.spatial_transform(frames)
+#         assert(frames.shape[2] == self.resolution[0] and frames.shape[3] == self.resolution[1]), f'frames={frames.shape}, self.resolution={self.resolution}'
+#         frames = frames.byte()
+#         # fps
+#         fps_clip = fps_ori // self.frame_stride
+#         if self.fps_max is not None and fps_clip > self.fps_max:
+#             fps_clip = self.fps_max
+        
+#         # caption index
+#         middle_idx = (rand_idx + self.video_length /2 )*fs
+#         big_cap_idx = (middle_idx // 64 +1) *64
+#         small_cap_idx = (middle_idx // 64) *64
+#         if big_cap_idx >= allf or ((big_cap_idx-middle_idx) >= (small_cap_idx-middle_idx)):
+#             cap_idx = small_cap_idx
+#         else:
+#             cap_idx = big_cap_idx
+#         # print(middle_idx, small_cap_idx, big_cap_idx,cap_idx)
+#         caption = sample[1][int(cap_idx//64)]
+
+#         frames = frames.permute(1,0,2,3)
+#         skeleton_final = torch.zeros_like(frames).byte()
+#         frames = (frames / 127.5 - 1.0)
+#         skeleton_final = (skeleton_final / 127.5 - 1.0)
+#         example = {'pixel_values': frames, 'sentence': caption, 'pose': skeleton_final}
+
+
+  
+#         return example
+    
+#     def __len__(self):
+#         return len(self.metadata)
+#         # return 1
+
+
+# if __name__ == '__main__':
+#     if True: # val
+#         hd_data = HDVila('/apdcephfs_cq3/share_1290939/0_public_datasets/hd-vila-100m','/apdcephfs_cq3/share_1290939/0_public_datasets/hd-vila-100m/caption_2048_val.csv')
+#     else:
+#         hd_data = HDVila('/apdcephfs_cq3/share_1290939/0_public_datasets/hd-vila-100m','/apdcephfs_cq3/share_1290939/0_public_datasets/hd-vila-100m/caption_rm2048_train.csv')
+#     print(len(hd_data))
+#     for i in range(len(hd_data)):
+#         # print(i) 
+#         hd_data[i]

FollowYourPose/followyourpose/models/attention.py

@@ -0,0 +1,375 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.models.attention import CrossAttention, FeedForward, AdaLayerNorm
+
+from einops import rearrange, repeat
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+
+        # Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
+        # Input
+        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length)
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                timestep=timestep,
+                video_length=video_length
+            )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+
+        output = hidden_states + residual
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+
+        # SC-Attn
+        self.attn1 = SparseCausalAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = CrossAttention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+
+        # Temp-Attn
+        self.attn_temp = CrossAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        # nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+        self.norm_temp = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+        self.conv_temporal = LoRALinearLayer(dim, dim, rank=160, stride=1)
+        
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        if not is_xformers_available():
+            print("Here is how to install it")
+            raise ModuleNotFoundError(
+                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                " xformers",
+                name="xformers",
+            )
+        elif not torch.cuda.is_available():
+            raise ValueError(
+                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
+                " available for GPU "
+            )
+        else:
+            try:
+                # Make sure we can run the memory efficient attention
+                _ = xformers.ops.memory_efficient_attention(
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                )
+            except Exception as e:
+                raise e
+            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            if self.attn2 is not None:
+                self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            # self.attn_temp._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None):
+        # SparseCausal-Attention
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
+        )
+
+        if self.only_cross_attention:
+            hidden_states = (
+                self.attn1(norm_hidden_states, encoder_hidden_states, attention_mask=attention_mask) + hidden_states
+            )
+        else:
+            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + hidden_states
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        d = hidden_states.shape[1]
+        hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=video_length)
+        norm_hidden_states = (
+            self.norm_temp(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_temp(hidden_states)
+        )
+        hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+        hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        
+        hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) c f", d=d, f=video_length)
+        hidden_states = self.conv_temporal(hidden_states)
+        hidden_states = rearrange(hidden_states, "(b d) c f -> (b f) d c", d=d, f=video_length)
+        
+        return hidden_states
+
+
+
+
+class LoRALinearLayer(nn.Module):
+    def __init__(self, in_features, out_features, rank=4, stride=1):
+        super().__init__()
+
+        if rank > min(in_features, out_features):
+            Warning(f"LoRA rank {rank} must be less or equal than {min(in_features, out_features)}, reset to {min(in_features, out_features)//2}")
+            rank = min(in_features, out_features)//2
+            
+
+        self.down = nn.Conv1d(in_features, rank, bias=False,
+                                kernel_size=3,
+                                stride = stride,
+                                padding=1,)
+        self.up = nn.Conv1d(rank, out_features, bias=False,
+                            kernel_size=3,
+                            padding=1,)
+
+        nn.init.normal_(self.down.weight, std=1 / rank)
+        # nn.init.zeros_(self.down.bias.data)
+        
+        nn.init.zeros_(self.up.weight)
+        # nn.init.zeros_(self.up.bias.data)
+        if stride > 1:
+            self.skip = nn.AvgPool1d(kernel_size=3, stride=2, padding=1)
+            
+    def forward(self, hidden_states):
+        orig_dtype = hidden_states.dtype
+        dtype = self.down.weight.dtype
+
+        down_hidden_states = self.down(hidden_states.to(dtype))
+        up_hidden_states = self.up(down_hidden_states)
+        if hasattr(self, 'skip'):
+            hidden_states=self.skip(hidden_states)
+        return up_hidden_states.to(orig_dtype)+hidden_states
+    
+
+
+
+class SparseCausalAttention(CrossAttention):
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, video_length=None):
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        encoder_hidden_states = encoder_hidden_states
+
+        if self.group_norm is not None:
+            hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = self.to_q(hidden_states)
+        dim = query.shape[-1]
+        query = self.reshape_heads_to_batch_dim(query)
+
+        if self.added_kv_proj_dim is not None:
+            raise NotImplementedError
+
+        encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        former_frame_index = torch.arange(video_length) - 1
+        former_frame_index[0] = 0
+
+        key = rearrange(key, "(b f) d c -> b f d c", f=video_length)
+        key = torch.cat([key[:, [0] * video_length], key[:, former_frame_index]], dim=2)
+        key = rearrange(key, "b f d c -> (b f) d c")
+
+        value = rearrange(value, "(b f) d c -> b f d c", f=video_length)
+        value = torch.cat([value[:, [0] * video_length], value[:, former_frame_index]], dim=2)
+        value = rearrange(value, "b f d c -> (b f) d c")
+
+        key = self.reshape_heads_to_batch_dim(key)
+        value = self.reshape_heads_to_batch_dim(value)
+
+        if attention_mask is not None:
+            if attention_mask.shape[-1] != query.shape[1]:
+                target_length = query.shape[1]
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+                attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)
+
+        # attention, what we cannot get enough of
+        if self._use_memory_efficient_attention_xformers:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+            # Some versions of xformers return output in fp32, cast it back to the dtype of the input
+            hidden_states = hidden_states.to(query.dtype)
+        else:
+            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
+                hidden_states = self._attention(query, key, value, attention_mask)
+            else:
+                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)
+
+        # linear proj
+        hidden_states = self.to_out[0](hidden_states)
+
+        # dropout
+        hidden_states = self.to_out[1](hidden_states)
+        return hidden_states

FollowYourPose/followyourpose/models/resnet.py

@@ -0,0 +1,209 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from einops import rearrange
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(self, channels, use_conv=False, use_conv_transpose=False, out_channels=None, name="conv"):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        conv = None
+        if use_conv_transpose:
+            raise NotImplementedError
+        elif use_conv:
+            conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+        if name == "conv":
+            self.conv = conv
+        else:
+            self.Conv2d_0 = conv
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest")
+        else:
+            hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        if self.use_conv:
+            if self.name == "conv":
+                hidden_states = self.conv(hidden_states)
+            else:
+                hidden_states = self.Conv2d_0(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(self, channels, use_conv=False, out_channels=None, padding=1, name="conv"):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            conv = InflatedConv3d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            raise NotImplementedError
+
+        if name == "conv":
+            self.Conv2d_0 = conv
+            self.conv = conv
+        elif name == "Conv2d_0":
+            self.conv = conv
+        else:
+            self.conv = conv
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        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.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+        self.conv1 = InflatedConv3d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+
+            self.time_emb_proj = torch.nn.Linear(temb_channels, time_emb_proj_out_channels)
+        else:
+            self.time_emb_proj = None
+
+        self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if non_linearity == "swish":
+            self.nonlinearity = lambda x: F.silu(x)
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = self.in_channels != self.out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))

FollowYourPose/followyourpose/models/unet.py

@@ -0,0 +1,571 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import os
+import json
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from .unet_blocks import (
+    CrossAttnDownBlock3D,
+    CrossAttnUpBlock3D,
+    DownBlock3D,
+    UNetMidBlock3DCrossAttn,
+    UpBlock3D,
+    get_down_block,
+    get_up_block,
+    conv_nd,
+    avg_pool_nd,
+)
+from .resnet import InflatedConv3d
+from einops import rearrange
+import sys
+sys.path.append('FollowYourPose')
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        mid_block_type: str = "UNetMidBlock3DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D"
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: int = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = InflatedConv3d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock3DCrossAttn":
+            self.mid_block = UNetMidBlock3DCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the videos
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=reversed_attention_head_dim[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)
+        self.conv_act = nn.SiLU()
+        self.conv_out = InflatedConv3d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+        
+        self.skeleton_adapter = Adapter(cin=int(3*64), channels=[320, 640, 1280, 1280][:4], nums_rb=2, ksize=1, sk=True, use_conv=False)
+        adapter_weight = torch.load('./FollowYourPose/checkpoints/t2iadapter_keypose_sd14v1.pth')
+        self.skeleton_adapter.load_state_dict(adapter_weight)
+        
+        
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
+
+        num_slicable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        skeleton: Optional[torch.FloatTensor] = None,
+        train_or_sample: str = 'train',
+    ) -> Union[UNet3DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        # pre-process
+        sample = self.conv_in(sample)
+
+        # down
+        down_block_res_samples = (sample,)
+        features_adapter = self.skeleton_adapter(skeleton)
+        
+        for idx, downsample_block in enumerate(self.down_blocks):
+            
+            if train_or_sample == 'train':
+                skeleton_feature = None
+            else:
+                skeleton_feature = features_adapter[idx]
+                
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    features_adapter=skeleton_feature
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb, features_adapter=skeleton_feature)
+
+            down_block_res_samples += res_samples
+
+        # mid
+        sample = self.mid_block(
+            sample, emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
+        )
+
+        # up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)
+
+    @classmethod
+    def from_pretrained_2d(cls, pretrained_model_path, subfolder=None):
+        if subfolder is not None:
+            pretrained_model_path = os.path.join(pretrained_model_path, subfolder)
+
+        config_file = os.path.join(pretrained_model_path, 'config.json')
+        if not os.path.isfile(config_file):
+            raise RuntimeError(f"{config_file} does not exist")
+        with open(config_file, "r") as f:
+            config = json.load(f)
+        config["_class_name"] = cls.__name__
+        config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D"
+        ]
+        config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D"
+        ]
+
+        from diffusers.utils import WEIGHTS_NAME
+        model = cls.from_config(config)
+        model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME)
+        if not os.path.isfile(model_file):
+            raise RuntimeError(f"{model_file} does not exist")
+        state_dict = torch.load(model_file, map_location="cpu")
+        for k, v in model.state_dict().items():
+            if '_temp.' in k:
+                state_dict.update({k: v})
+        model.load_state_dict(state_dict, strict=False)
+
+        return model
+    
+    
+    
+
+class Adapter(nn.Module):
+    def __init__(self, channels=[320, 640, 1280, 1280], nums_rb=3, cin=64, ksize=3, sk=False, use_conv=True):
+        super(Adapter, self).__init__()
+        self.unshuffle = nn.PixelUnshuffle(8)
+        self.channels = channels
+        self.nums_rb = nums_rb
+        self.body = []
+        for i in range(len(channels)):
+            for j in range(nums_rb):
+                if (i!=0) and (j==0):
+                    self.body.append(ResnetBlock(channels[i-1], channels[i], down=True, ksize=ksize, sk=sk, use_conv=use_conv))
+                else:
+                    self.body.append(ResnetBlock(channels[i], channels[i], down=False, ksize=ksize, sk=sk, use_conv=use_conv))
+        self.body = nn.ModuleList(self.body)
+        self.conv_in = nn.Conv2d(cin,channels[0], 3, 1, 1)
+
+    def forward(self, x):
+        b, t, c, h, w = x.shape
+        x = rearrange(x, 'b t c h w -> (b t) c h w')
+        # unshuffle
+        x = self.unshuffle(x)
+        # extract features
+        features = []
+        x = self.conv_in(x)
+        for i in range(len(self.channels)):
+            for j in range(self.nums_rb):
+                idx = i*self.nums_rb +j
+                x = self.body[idx](x)
+            features.append(x)
+
+        features = [ rearrange(fn, '(b t) c h w -> b c t h w', b=b, t=t) for fn in features]
+        return features
+    
+    
+
+class ResnetBlock(nn.Module):
+    def __init__(self, in_c, out_c, down, ksize=3, sk=False, use_conv=True):
+        super().__init__()
+        ps = ksize//2
+        if in_c != out_c or sk==False:
+            self.in_conv = nn.Conv2d(in_c, out_c, ksize, 1, ps)
+        else:
+            # print('n_in')
+            self.in_conv = None
+        self.block1 = nn.Conv2d(out_c, out_c, 3, 1, 1)
+        self.act = nn.ReLU()
+        self.block2 = nn.Conv2d(out_c, out_c, ksize, 1, ps)
+        if sk==False:
+            self.skep = nn.Conv2d(in_c, out_c, ksize, 1, ps)
+        else:
+            self.skep = None
+
+        self.down = down
+        if self.down == True:
+            self.down_opt = Downsample(in_c, use_conv=use_conv)
+
+    def forward(self, x):
+        if self.down == True:
+            x = self.down_opt(x)
+        if self.in_conv is not None: # edit
+            x = self.in_conv(x)
+
+        h = self.block1(x)
+        h = self.act(h)
+        h = self.block2(h)
+        if self.skep is not None:
+            return h + self.skep(x)
+        else:
+            return h + 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)

FollowYourPose/followyourpose/models/unet_blocks.py

@@ -0,0 +1,631 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import torch
+from torch import nn
+
+from .attention import Transformer3DModel
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+):
+    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+    if down_block_type == "DownBlock3D":
+        return DownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "CrossAttnDownBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock3D")
+        return CrossAttnDownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+):
+    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    if up_block_type == "UpBlock3D":
+        return UpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "CrossAttnUpBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock3D")
+        return CrossAttnUpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for _ in range(num_layers):
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    upcast_attention=upcast_attention,
+                )
+            )
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                )
+            )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None, features_adapter=None):
+        output_states = ()
+        idx = 1
+        for resnet, attn in zip(self.resnets, self.attentions):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
+                if ((idx+1)%3 == 0) and features_adapter is not None and len(features_adapter):
+                    hidden_states = hidden_states + features_adapter
+                idx = idx + 1
+                
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_downsample=True,
+        downsample_padding=1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None, features_adapter=None):
+        output_states = ()
+        idx = 1
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                if ((idx+1)%3 == 0) and features_adapter is not None and len(features_adapter):
+                    hidden_states = hidden_states + features_adapter
+                idx = idx + 1
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample3D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        encoder_hidden_states=None,
+        upsample_size=None,
+        attention_mask=None,
+    ):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_upsample=True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample3D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+
+
+def conv_nd(dims, in_channels, out_channels, kernel_size, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(in_channels, out_channels, kernel_size, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(in_channels, out_channels, kernel_size, **kwargs)
+    elif dims == 3:
+        if isinstance(kernel_size, int):
+            kernel_size = (1, *((kernel_size,) * 2))
+        if 'stride' in kwargs.keys():
+            if isinstance(kwargs['stride'], int):
+                kwargs['stride'] = (1, *((kwargs['stride'],) * 2))
+        if 'padding' in kwargs.keys():
+            if isinstance(kwargs['padding'], int):
+                kwargs['padding'] = (0, *((kwargs['padding'],) * 2))
+        return nn.Conv3d(in_channels, out_channels, kernel_size, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+
+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}")

FollowYourPose/followyourpose/pipelines/pipeline_followyourpose.py

@@ -0,0 +1,442 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
+
+import inspect
+from typing import Callable, List, Optional, Union
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+
+from diffusers.utils import is_accelerate_available
+from packaging import version
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from diffusers.utils import deprecate, logging, BaseOutput
+
+from einops import rearrange
+
+from ..models.unet import UNet3DConditionModel
+from torchvision import transforms
+import torchvision.transforms._transforms_video as transforms_video
+import decord
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class FollowYourPosePipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+
+class FollowYourPosePipeline(DiffusionPipeline):
+    _optional_components = []
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet3DConditionModel,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+        ],
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+            )
+            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["clip_sample"] = False
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+            version.parse(unet.config._diffusers_version).base_version
+        ) < version.parse("0.9.0.dev0")
+        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+            deprecation_message = (
+                "The configuration file of the unet has set the default `sample_size` to smaller than"
+                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+                " in the config might lead to incorrect results in future versions. If you have downloaded this"
+                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+                " the `unet/config.json` file"
+            )
+            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(unet.config)
+            new_config["sample_size"] = 64
+            unet._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+    def enable_vae_slicing(self):
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_prompt(self, prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt):
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+
+        if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = text_embeddings[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = uncond_embeddings[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        video = self.vae.decode(latents).sample
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(self, prompt, height, width, callback_steps):
+        if not isinstance(prompt, str) and not isinstance(prompt, list):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    def prepare_latents(self, batch_size, num_channels_latents, video_length, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, video_length, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            rand_device = "cpu" if device.type == "mps" else device
+
+            if isinstance(generator, list):
+                shape = (1,) + shape[1:]
+                latents = [
+                    torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype)
+                    for i in range(batch_size)
+                ]
+                latents = torch.cat(latents, dim=0).to(device)
+            else:
+                latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype).to(device)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        video_length: Optional[int],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        skeleton_path: Optional[str] = None,
+        **kwargs,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, callback_steps)
+
+        # Define call parameters
+        batch_size = 1 if isinstance(prompt, str) else len(prompt)
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Encode input prompt
+        text_embeddings = self._encode_prompt(
+            prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
+        )
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # Prepare latent variables
+        num_channels_latents = self.unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            video_length,
+            height,
+            width,
+            text_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+        latents_dtype = latents.dtype
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        
+        skeleton, save_skeleton = self.get_skeleton(skeleton_path)
+        skeleton = skeleton.to(latents.device).repeat(2,1,1,1,1)
+        
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings, skeleton=skeleton, train_or_sample='sample').sample.to(dtype=latents_dtype)
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        # Post-processing
+        video = self.decode_latents(latents)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            video = torch.from_numpy(video)
+            video_skeleton = rearrange(save_skeleton, 'b t h w c -> b c t h w')
+            video = torch.cat([video_skeleton,video],dim=-1)
+
+        if not return_dict:
+            return video
+
+        return FollowYourPosePipelineOutput(videos=video)
+    
+    
+    @torch.no_grad()
+    def get_skeleton(self,skeleton_path):
+        skeleton_start_end = list(range(0, 120, 5))
+        self_transform = transforms.Compose([transforms.Resize(512),
+                                        transforms_video.CenterCropVideo(512)])
+        
+        vr_skeleton = decord.VideoReader(skeleton_path)
+        # sample frames
+
+        # for start_end in skeleton_start_end:
+        skeleton = vr_skeleton.get_batch(skeleton_start_end)
+        if not isinstance(skeleton,torch.Tensor):
+            skeleton = torch.from_numpy(skeleton.asnumpy()).float()
+        skeleton_video_trans = self_transform(skeleton.permute(3, 0, 1, 2))
+        skeleton_final = skeleton_video_trans.permute(1, 2, 3, 0)
+        # import pdb;pdb.set_trace()
+
+        skeleton_video = (skeleton_final / 255).unsqueeze(0)
+        save_skeleton = skeleton_video.clone().detach()
+        skeleton_video = rearrange(skeleton_video, 'b t h w c -> b t c h w')
+        # skeleton_video = skeleton_video.to(model.device)
+        
+        return skeleton_video, save_skeleton

FollowYourPose/followyourpose/util.py

@@ -0,0 +1,84 @@
+import os
+import imageio
+import numpy as np
+from typing import Union
+
+import torch
+import torchvision
+
+from tqdm import tqdm
+from einops import rearrange
+
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=4, fps=8):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    outputs = []
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        outputs.append(x)
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    imageio.mimsave(path, outputs, fps=fps)
+
+
+# DDIM Inversion
+@torch.no_grad()
+def init_prompt(prompt, pipeline):
+    uncond_input = pipeline.tokenizer(
+        [""], padding="max_length", max_length=pipeline.tokenizer.model_max_length,
+        return_tensors="pt"
+    )
+    uncond_embeddings = pipeline.text_encoder(uncond_input.input_ids.to(pipeline.device))[0]
+    text_input = pipeline.tokenizer(
+        [prompt],
+        padding="max_length",
+        max_length=pipeline.tokenizer.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    text_embeddings = pipeline.text_encoder(text_input.input_ids.to(pipeline.device))[0]
+    context = torch.cat([uncond_embeddings, text_embeddings])
+
+    return context
+
+
+def next_step(model_output: Union[torch.FloatTensor, np.ndarray], timestep: int,
+              sample: Union[torch.FloatTensor, np.ndarray], ddim_scheduler):
+    timestep, next_timestep = min(
+        timestep - ddim_scheduler.config.num_train_timesteps // ddim_scheduler.num_inference_steps, 999), timestep
+    alpha_prod_t = ddim_scheduler.alphas_cumprod[timestep] if timestep >= 0 else ddim_scheduler.final_alpha_cumprod
+    alpha_prod_t_next = ddim_scheduler.alphas_cumprod[next_timestep]
+    beta_prod_t = 1 - alpha_prod_t
+    next_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
+    next_sample_direction = (1 - alpha_prod_t_next) ** 0.5 * model_output
+    next_sample = alpha_prod_t_next ** 0.5 * next_original_sample + next_sample_direction
+    return next_sample
+
+
+def get_noise_pred_single(latents, t, context, unet):
+    noise_pred = unet(latents, t, encoder_hidden_states=context)["sample"]
+    return noise_pred
+
+
+@torch.no_grad()
+def ddim_loop(pipeline, ddim_scheduler, latent, num_inv_steps, prompt):
+    context = init_prompt(prompt, pipeline)
+    uncond_embeddings, cond_embeddings = context.chunk(2)
+    all_latent = [latent]
+    latent = latent.clone().detach()
+    for i in tqdm(range(num_inv_steps)):
+        t = ddim_scheduler.timesteps[len(ddim_scheduler.timesteps) - i - 1]
+        noise_pred = get_noise_pred_single(latent, t, cond_embeddings, pipeline.unet)
+        latent = next_step(noise_pred, t, latent, ddim_scheduler)
+        all_latent.append(latent)
+    return all_latent
+
+
+@torch.no_grad()
+def ddim_inversion(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt=""):
+    ddim_latents = ddim_loop(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt)
+    return ddim_latents

FollowYourPose/test_followyourpose.py

@@ -0,0 +1,189 @@
+import argparse
+import datetime
+import logging
+import inspect
+import math
+import os
+from typing import Dict, Optional, Tuple
+from omegaconf import OmegaConf
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+
+import diffusers
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import set_seed
+from diffusers import AutoencoderKL, DDPMScheduler, DDIMScheduler
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version
+from diffusers.utils.import_utils import is_xformers_available
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTokenizer
+
+import sys
+sys.path.append('FollowYourPose')
+from followyourpose.models.unet import UNet3DConditionModel
+from followyourpose.pipelines.pipeline_followyourpose import FollowYourPosePipeline
+from followyourpose.util import save_videos_grid, ddim_inversion
+from einops import rearrange
+
+check_min_version("0.10.0.dev0")
+
+logger = get_logger(__name__, log_level="INFO")
+
+
+def collate_fn(examples):
+    """Concat a batch of sampled image in dataloader
+    """
+    batch = {
+        "prompt_ids": torch.cat([example["prompt_ids"] for example in examples], dim=0),
+        "images": torch.stack([example["images"] for example in examples]),
+    }
+    return batch
+
+
+
+def test(
+    pretrained_model_path: str,
+    output_dir: str,
+    validation_data: Dict,
+    validation_steps: int = 100,
+    train_batch_size: int = 1,
+    gradient_accumulation_steps: int = 1,
+    gradient_checkpointing: bool = True,
+    resume_from_checkpoint: Optional[str] = None,
+    mixed_precision: Optional[str] = "fp16",
+    enable_xformers_memory_efficient_attention: bool = True,
+    seed: Optional[int] = None,
+    skeleton_path: Optional[str] = None,
+):
+    *_, config = inspect.getargvalues(inspect.currentframe())
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=gradient_accumulation_steps,
+        mixed_precision=mixed_precision,
+    )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state, main_process_only=False)
+    if accelerator.is_local_main_process:
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    # If passed along, set the training seed now.
+    if seed is not None:
+        set_seed(seed)
+
+    # Handle the output folder creation
+    if accelerator.is_main_process:
+        
+        os.makedirs(output_dir, exist_ok=True)
+        os.makedirs(f"{output_dir}/samples", exist_ok=True)
+        os.makedirs(f"{output_dir}/inv_latents", exist_ok=True)
+        OmegaConf.save(config, os.path.join(output_dir, 'config.yaml'))
+
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
+    tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder")
+    vae = AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae")
+    unet = UNet3DConditionModel.from_pretrained_2d(pretrained_model_path, subfolder="unet")
+
+    # Freeze vae and text_encoder
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+
+    unet.requires_grad_(False)
+    # for name, module in unet.named_modules():
+    #     if name.endswith(tuple(trainable_modules)):
+    #         for params in module.parameters():
+    #             params.requires_grad = True
+
+    if enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            unet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    if gradient_checkpointing:
+        unet.enable_gradient_checkpointing()
+
+
+    # Get the validation pipeline
+    validation_pipeline = FollowYourPosePipeline(
+        vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet,
+        scheduler=DDIMScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
+    )
+    validation_pipeline.enable_vae_slicing()
+    ddim_inv_scheduler = DDIMScheduler.from_pretrained(pretrained_model_path, subfolder='scheduler')
+    ddim_inv_scheduler.set_timesteps(validation_data.num_inv_steps)
+
+    unet = accelerator.prepare(unet)
+    # For mixed precision training we cast the text_encoder and vae weights to half-precision
+    # as these models are only used for inference, keeping weights in full precision is not required.
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    # Move text_encode and vae to gpu and cast to weight_dtype
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        accelerator.init_trackers("text2video-fine-tune")
+
+    global_step = 0
+    first_epoch = 0
+
+    # Potentially load in the weights and states from a previous save
+    load_path = None
+    if resume_from_checkpoint:
+        if resume_from_checkpoint != "latest":
+  
+            load_path = resume_from_checkpoint
+            output_dir = os.path.abspath(os.path.join(resume_from_checkpoint, ".."))
+        accelerator.print(f"load from checkpoint {load_path}")
+        accelerator.load_state(load_path)
+
+        global_step = int(load_path.split("-")[-1])
+
+                
+    if accelerator.is_main_process:
+        samples = []
+        generator = torch.Generator(device=accelerator.device)
+        generator.manual_seed(seed)
+
+        ddim_inv_latent = None
+        from datetime import datetime
+        now = str(datetime.now())
+        print(now)
+        for idx, prompt in enumerate(validation_data.prompts):
+            sample = validation_pipeline(prompt, generator=generator, latents=ddim_inv_latent,
+                                        skeleton_path=skeleton_path,
+                                        **validation_data).videos
+            save_videos_grid(sample, f"{output_dir}/inference/sample-{global_step}-{str(seed)}-{now}/{prompt}.gif")
+            samples.append(sample)
+        samples = torch.concat(samples)
+        save_path = f"{output_dir}/inference/sample-{global_step}-{str(seed)}-{now}.mp4"
+        save_videos_grid(samples, save_path)
+        logger.info(f"Saved samples to {save_path}")
+        
+    return save_path
+