weifeng.genius commited on
Commit
3f4baa8
1 Parent(s): a2f07f6

first init

Browse files
Files changed (28) hide show
  1. app.py +202 -0
  2. model/annotator/canny/__init__.py +6 -0
  3. model/annotator/hed/__init__.py +133 -0
  4. model/annotator/hed/__pycache__/__init__.cpython-39.pyc +0 -0
  5. model/annotator/util.py +38 -0
  6. model/video_diffusion/__init__.py +0 -0
  7. model/video_diffusion/__pycache__/__init__.cpython-39.pyc +0 -0
  8. model/video_diffusion/models/__init__.py +0 -0
  9. model/video_diffusion/models/__pycache__/__init__.cpython-39.pyc +0 -0
  10. model/video_diffusion/models/__pycache__/attention.cpython-39.pyc +0 -0
  11. model/video_diffusion/models/__pycache__/controlnet3d.cpython-39.pyc +0 -0
  12. model/video_diffusion/models/__pycache__/resnet.cpython-39.pyc +0 -0
  13. model/video_diffusion/models/__pycache__/unet_3d_blocks.cpython-39.pyc +0 -0
  14. model/video_diffusion/models/__pycache__/unet_3d_blocks_control.cpython-39.pyc +0 -0
  15. model/video_diffusion/models/__pycache__/unet_3d_condition.cpython-39.pyc +0 -0
  16. model/video_diffusion/models/attention.py +454 -0
  17. model/video_diffusion/models/controlnet3d.py +580 -0
  18. model/video_diffusion/models/resnet.py +486 -0
  19. model/video_diffusion/models/unet_3d_blocks.py +622 -0
  20. model/video_diffusion/models/unet_3d_blocks_control.py +116 -0
  21. model/video_diffusion/models/unet_3d_condition.py +571 -0
  22. model/video_diffusion/pipelines/__init__.py +0 -0
  23. model/video_diffusion/pipelines/__pycache__/__init__.cpython-39.pyc +0 -0
  24. model/video_diffusion/pipelines/__pycache__/pipeline_st_stable_diffusion.cpython-39.pyc +0 -0
  25. model/video_diffusion/pipelines/__pycache__/pipeline_stable_diffusion_controlnet3d.cpython-39.pyc +0 -0
  26. model/video_diffusion/pipelines/pipeline_st_stable_diffusion.py +618 -0
  27. model/video_diffusion/pipelines/pipeline_stable_diffusion_controlnet3d.py +482 -0
  28. requirements.txt +8 -0
app.py ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from model.video_diffusion.models.controlnet3d import ControlNet3DModel
2
+ from model.video_diffusion.models.unet_3d_condition import UNetPseudo3DConditionModel
3
+ from model.video_diffusion.pipelines.pipeline_stable_diffusion_controlnet3d import Controlnet3DStableDiffusionPipeline
4
+ from transformers import DPTForDepthEstimation
5
+ from model.annotator.hed import HEDNetwork
6
+ import torch
7
+ from einops import rearrange,repeat
8
+ import imageio
9
+ import numpy as np
10
+ import cv2
11
+ import torch.nn.functional as F
12
+ from PIL import Image
13
+ import argparse
14
+ import tempfile
15
+ import os
16
+ import gradio as gr
17
+
18
+
19
+ control_mode = 'depth'
20
+ control_net_path = f"wf-genius/controlavideo-{control_mode}"
21
+ unet = UNetPseudo3DConditionModel.from_pretrained(control_net_path,
22
+ torch_dtype = torch.float16,
23
+ subfolder='unet',
24
+ ).to("cuda")
25
+ controlnet = ControlNet3DModel.from_pretrained(control_net_path,
26
+ torch_dtype = torch.float16,
27
+ subfolder='controlnet',
28
+ ).to("cuda")
29
+
30
+ if control_mode == 'depth':
31
+ annotator_model = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
32
+ elif control_mode == 'canny':
33
+ annotator_model = None
34
+ elif control_mode == 'hed':
35
+ # firstly download from https://huggingface.co/wf-genius/controlavideo-hed/resolve/main/hed-network.pth
36
+ annotator_model = HEDNetwork('hed-network.pth').to("cuda")
37
+
38
+ video_controlnet_pipe = Controlnet3DStableDiffusionPipeline.from_pretrained(control_net_path, unet=unet,
39
+ controlnet=controlnet, annotator_model=annotator_model,
40
+ torch_dtype = torch.float16,
41
+ ).to("cuda")
42
+
43
+
44
+ def to_video(frames, fps: int) -> str:
45
+ out_file = tempfile.NamedTemporaryFile(suffix='.mp4', delete=False)
46
+ writer = imageio.get_writer(out_file.name, format='FFMPEG', fps=fps)
47
+ for frame in frames:
48
+ writer.append_data(np.array(frame))
49
+ writer.close()
50
+ return out_file.name
51
+
52
+ def inference(input_video,
53
+ prompt,
54
+ seed,
55
+ num_inference_steps,
56
+ guidance_scale,
57
+ sampling_rate,
58
+ video_scale,
59
+ init_noise_thres,
60
+ each_sample_frame,
61
+ iter_times,
62
+ h,
63
+ w,
64
+ ):
65
+ num_sample_frames = iter_times * each_sample_frame
66
+ testing_prompt = [prompt]
67
+ np_frames, fps_vid = Controlnet3DStableDiffusionPipeline.get_frames_preprocess(input_video, num_frames=num_sample_frames, sampling_rate=sampling_rate, return_np=True)
68
+ if control_mode == 'depth':
69
+ frames = torch.from_numpy(np_frames).div(255) * 2 - 1
70
+ frames = rearrange(frames, "f h w c -> c f h w").unsqueeze(0)
71
+ frames = rearrange(frames, 'b c f h w -> (b f) c h w')
72
+ control_maps = video_controlnet_pipe.get_depth_map(frames, h, w, return_standard_norm=False) # (b f) 1 h w
73
+ elif control_mode == 'canny':
74
+ control_maps = np.stack([cv2.Canny(inp, 100, 200) for inp in np_frames])
75
+ control_maps = repeat(control_maps, 'f h w -> f c h w',c=1)
76
+ control_maps = torch.from_numpy(control_maps).div(255) # 0~1
77
+ elif control_mode == 'hed':
78
+ control_maps = np.stack([video_controlnet_pipe.get_hed_map(inp) for inp in np_frames])
79
+ control_maps = repeat(control_maps, 'f h w -> f c h w',c=1)
80
+ control_maps = torch.from_numpy(control_maps).div(255) # 0~1
81
+ control_maps = control_maps.to(dtype=controlnet.dtype, device=controlnet.device)
82
+ control_maps = F.interpolate(control_maps, size=(h,w), mode='bilinear', align_corners=False)
83
+ control_maps = rearrange(control_maps, "(b f) c h w -> b c f h w", f=num_sample_frames)
84
+ if control_maps.shape[1] == 1:
85
+ control_maps = repeat(control_maps, 'b c f h w -> b (n c) f h w', n=3)
86
+
87
+ frames = torch.from_numpy(np_frames).div(255)
88
+ frames = rearrange(frames, 'f h w c -> f c h w')
89
+ v2v_input_frames = torch.nn.functional.interpolate(
90
+ frames,
91
+ size=(h, w),
92
+ mode="bicubic",
93
+ antialias=True,
94
+ )
95
+ v2v_input_frames = rearrange(v2v_input_frames, '(b f) c h w -> b c f h w ', f=num_sample_frames)
96
+
97
+ out = []
98
+ for i in range(num_sample_frames//each_sample_frame):
99
+ out1 = video_controlnet_pipe(
100
+ # controlnet_hint= control_maps[:,:,:each_sample_frame,:,:],
101
+ # images= v2v_input_frames[:,:,:each_sample_frame,:,:],
102
+ controlnet_hint=control_maps[:,:,i*each_sample_frame-1:(i+1)*each_sample_frame-1,:,:] if i>0 else control_maps[:,:,:each_sample_frame,:,:],
103
+ images=v2v_input_frames[:,:,i*each_sample_frame-1:(i+1)*each_sample_frame-1,:,:] if i>0 else v2v_input_frames[:,:,:each_sample_frame,:,:],
104
+ first_frame_output=out[-1] if i>0 else None,
105
+ prompt=testing_prompt,
106
+ num_inference_steps=num_inference_steps,
107
+ width=w,
108
+ height=h,
109
+ guidance_scale=guidance_scale,
110
+ generator=[torch.Generator(device="cuda").manual_seed(seed)],
111
+ video_scale = video_scale,
112
+ init_noise_by_residual_thres = init_noise_thres, # residual-based init. larger thres ==> more smooth.
113
+ controlnet_conditioning_scale=1.0,
114
+ fix_first_frame=True,
115
+ in_domain=True,
116
+ )
117
+ out1 = out1.images[0]
118
+ if len(out1) > 1:
119
+ out1 = out1[1:] # drop the first frame
120
+ out.extend(out1)
121
+
122
+ return to_video(out, 8)
123
+
124
+
125
+ examples = [
126
+ ["__assets__/depth_videos_depth/girl_dancing.mp4",
127
+ "A stormtrooper, masterpiece, a high-quality, detailed, and professional photo"],
128
+ ]
129
+
130
+ def preview_inference(
131
+ input_video,
132
+ prompt, seed,
133
+ num_inference_steps, guidance_scale,
134
+ sampling_rate, video_scale, init_noise_thres,
135
+ each_sample_frame,iter_times, h, w,
136
+ ):
137
+ return inference(input_video,
138
+ prompt, seed,
139
+ num_inference_steps, guidance_scale,
140
+ sampling_rate, 0.0, 0.0, 1, 1, h, w,)
141
+
142
+ if __name__ == '__main__':
143
+ with gr.Blocks() as demo:
144
+ with gr.Row():
145
+ with gr.Column():
146
+ input_video = gr.Video(
147
+ label="Input Video", source='upload', format="mp4", visible=True)
148
+ with gr.Column():
149
+ init_noise_thres = gr.Slider(0, 1, value=0.1, step=0.1, label="init_noise_thress")
150
+ each_sample_frame = gr.Slider(6, 16, value=8, step=1, label="each_sample_frame")
151
+ iter_times = gr.Slider(1, 4, value=1, step=1, label="iter_times")
152
+ sampling_rate = gr.Slider(1, 8, value=3, step=1, label="sampling_rate")
153
+ h = gr.Slider(256, 768, value=512, step=64, label="height")
154
+ w = gr.Slider(256, 768, value=512, step=64, label="width")
155
+ with gr.Column():
156
+ seed = gr.Slider(0, 6666, value=1, step=1, label="seed")
157
+ num_inference_steps = gr.Slider(5, 50, value=20, step=1, label="num_inference_steps")
158
+ guidance_scale = gr.Slider(1, 20, value=7.5, step=0.5, label="guidance_scale")
159
+ video_scale = gr.Slider(0, 2.5, value=1.5, step=0.1, label="video_scale")
160
+ prompt = gr.Textbox(label='Prompt')
161
+ # preview_button = gr.Button('Preview')
162
+ run_button = gr.Button('Generate Video')
163
+
164
+ with gr.Column():
165
+ result = gr.Video(label="Generated Video")
166
+
167
+ inputs = [
168
+ input_video,
169
+ prompt,
170
+ seed,
171
+ num_inference_steps,
172
+ guidance_scale,
173
+ sampling_rate,
174
+ video_scale,
175
+ init_noise_thres,
176
+ each_sample_frame,
177
+ iter_times,
178
+ h,
179
+ w,
180
+ ]
181
+
182
+ gr.Examples(examples=examples,
183
+ inputs=inputs,
184
+ outputs=result,
185
+ fn=inference,
186
+ cache_examples=False,
187
+ run_on_click=False,
188
+ )
189
+
190
+ run_button.click(fn=inference,
191
+ inputs=inputs,
192
+ outputs=result,)
193
+ # preview_button.click(fn=preview_inference,
194
+ # inputs=inputs,
195
+ # outputs=result,)
196
+
197
+ demo.launch(server_name="0.0.0.0", server_port=7860)
198
+
199
+
200
+ # TODO
201
+ # 1. preview
202
+ # 2. params
model/annotator/canny/__init__.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ import cv2
2
+
3
+
4
+ class CannyDetector:
5
+ def __call__(self, img, low_threshold, high_threshold):
6
+ return cv2.Canny(img, low_threshold, high_threshold)
model/annotator/hed/__init__.py ADDED
@@ -0,0 +1,133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import cv2
3
+ import os
4
+ import torch
5
+ from einops import rearrange
6
+
7
+
8
+ class HEDNetwork(torch.nn.Module):
9
+ def __init__(self, model_path):
10
+ super().__init__()
11
+
12
+ self.netVggOne = torch.nn.Sequential(
13
+ torch.nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1),
14
+ torch.nn.ReLU(inplace=False),
15
+ torch.nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1),
16
+ torch.nn.ReLU(inplace=False)
17
+ )
18
+
19
+ self.netVggTwo = torch.nn.Sequential(
20
+ torch.nn.MaxPool2d(kernel_size=2, stride=2),
21
+ torch.nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1),
22
+ torch.nn.ReLU(inplace=False),
23
+ torch.nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1),
24
+ torch.nn.ReLU(inplace=False)
25
+ )
26
+
27
+ self.netVggThr = torch.nn.Sequential(
28
+ torch.nn.MaxPool2d(kernel_size=2, stride=2),
29
+ torch.nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1),
30
+ torch.nn.ReLU(inplace=False),
31
+ torch.nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1),
32
+ torch.nn.ReLU(inplace=False),
33
+ torch.nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1),
34
+ torch.nn.ReLU(inplace=False)
35
+ )
36
+
37
+ self.netVggFou = torch.nn.Sequential(
38
+ torch.nn.MaxPool2d(kernel_size=2, stride=2),
39
+ torch.nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1),
40
+ torch.nn.ReLU(inplace=False),
41
+ torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
42
+ torch.nn.ReLU(inplace=False),
43
+ torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
44
+ torch.nn.ReLU(inplace=False)
45
+ )
46
+
47
+ self.netVggFiv = torch.nn.Sequential(
48
+ torch.nn.MaxPool2d(kernel_size=2, stride=2),
49
+ torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
50
+ torch.nn.ReLU(inplace=False),
51
+ torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
52
+ torch.nn.ReLU(inplace=False),
53
+ torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
54
+ torch.nn.ReLU(inplace=False)
55
+ )
56
+
57
+ self.netScoreOne = torch.nn.Conv2d(in_channels=64, out_channels=1, kernel_size=1, stride=1, padding=0)
58
+ self.netScoreTwo = torch.nn.Conv2d(in_channels=128, out_channels=1, kernel_size=1, stride=1, padding=0)
59
+ self.netScoreThr = torch.nn.Conv2d(in_channels=256, out_channels=1, kernel_size=1, stride=1, padding=0)
60
+ self.netScoreFou = torch.nn.Conv2d(in_channels=512, out_channels=1, kernel_size=1, stride=1, padding=0)
61
+ self.netScoreFiv = torch.nn.Conv2d(in_channels=512, out_channels=1, kernel_size=1, stride=1, padding=0)
62
+
63
+ self.netCombine = torch.nn.Sequential(
64
+ torch.nn.Conv2d(in_channels=5, out_channels=1, kernel_size=1, stride=1, padding=0),
65
+ torch.nn.Sigmoid()
66
+ )
67
+
68
+ self.load_state_dict({strKey.replace('module', 'net'): tenWeight for strKey, tenWeight in torch.load(model_path).items()})
69
+
70
+ def forward(self, tenInput):
71
+ tenInput = tenInput * 255.0
72
+ tenInput = tenInput - torch.tensor(data=[104.00698793, 116.66876762, 122.67891434], dtype=tenInput.dtype, device=tenInput.device).view(1, 3, 1, 1)
73
+
74
+ tenVggOne = self.netVggOne(tenInput)
75
+ tenVggTwo = self.netVggTwo(tenVggOne)
76
+ tenVggThr = self.netVggThr(tenVggTwo)
77
+ tenVggFou = self.netVggFou(tenVggThr)
78
+ tenVggFiv = self.netVggFiv(tenVggFou)
79
+
80
+ tenScoreOne = self.netScoreOne(tenVggOne)
81
+ tenScoreTwo = self.netScoreTwo(tenVggTwo)
82
+ tenScoreThr = self.netScoreThr(tenVggThr)
83
+ tenScoreFou = self.netScoreFou(tenVggFou)
84
+ tenScoreFiv = self.netScoreFiv(tenVggFiv)
85
+
86
+ tenScoreOne = torch.nn.functional.interpolate(input=tenScoreOne, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
87
+ tenScoreTwo = torch.nn.functional.interpolate(input=tenScoreTwo, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
88
+ tenScoreThr = torch.nn.functional.interpolate(input=tenScoreThr, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
89
+ tenScoreFou = torch.nn.functional.interpolate(input=tenScoreFou, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
90
+ tenScoreFiv = torch.nn.functional.interpolate(input=tenScoreFiv, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
91
+
92
+ return self.netCombine(torch.cat([ tenScoreOne, tenScoreTwo, tenScoreThr, tenScoreFou, tenScoreFiv ], 1))
93
+
94
+
95
+ class HEDdetector:
96
+ def __init__(self, network ):
97
+ self.netNetwork = network
98
+
99
+ def __call__(self, input_image):
100
+ if isinstance(input_image, torch.Tensor):
101
+ # 输入的就是 b c h w的tensor 范围是-1~1,需要转换为0~1
102
+ input_image = (input_image + 1) / 2
103
+ input_image = input_image.float().cuda()
104
+ edge = self.netNetwork(input_image) # 范围也是0~1, 不用转了直接用
105
+ return edge
106
+ else:
107
+ assert input_image.ndim == 3
108
+ input_image = input_image[:, :, ::-1].copy()
109
+ with torch.no_grad():
110
+ image_hed = torch.from_numpy(input_image).float().cuda()
111
+ image_hed = image_hed / 255.0
112
+ image_hed = rearrange(image_hed, 'h w c -> 1 c h w')
113
+ edge = self.netNetwork(image_hed)[0]
114
+ edge = (edge.cpu().numpy() * 255.0).clip(0, 255).astype(np.uint8)
115
+ return edge[0]
116
+
117
+
118
+ def nms(x, t, s):
119
+ x = cv2.GaussianBlur(x.astype(np.float32), (0, 0), s)
120
+
121
+ f1 = np.array([[0, 0, 0], [1, 1, 1], [0, 0, 0]], dtype=np.uint8)
122
+ f2 = np.array([[0, 1, 0], [0, 1, 0], [0, 1, 0]], dtype=np.uint8)
123
+ f3 = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=np.uint8)
124
+ f4 = np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]], dtype=np.uint8)
125
+
126
+ y = np.zeros_like(x)
127
+
128
+ for f in [f1, f2, f3, f4]:
129
+ np.putmask(y, cv2.dilate(x, kernel=f) == x, x)
130
+
131
+ z = np.zeros_like(y, dtype=np.uint8)
132
+ z[y > t] = 255
133
+ return z
model/annotator/hed/__pycache__/__init__.cpython-39.pyc ADDED
Binary file (4.31 kB). View file
 
model/annotator/util.py ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import cv2
3
+ import os
4
+
5
+
6
+ annotator_ckpts_path = os.path.join(os.path.dirname(__file__), 'ckpts')
7
+
8
+
9
+ def HWC3(x):
10
+ assert x.dtype == np.uint8
11
+ if x.ndim == 2:
12
+ x = x[:, :, None]
13
+ assert x.ndim == 3
14
+ H, W, C = x.shape
15
+ assert C == 1 or C == 3 or C == 4
16
+ if C == 3:
17
+ return x
18
+ if C == 1:
19
+ return np.concatenate([x, x, x], axis=2)
20
+ if C == 4:
21
+ color = x[:, :, 0:3].astype(np.float32)
22
+ alpha = x[:, :, 3:4].astype(np.float32) / 255.0
23
+ y = color * alpha + 255.0 * (1.0 - alpha)
24
+ y = y.clip(0, 255).astype(np.uint8)
25
+ return y
26
+
27
+
28
+ def resize_image(input_image, resolution):
29
+ H, W, C = input_image.shape
30
+ H = float(H)
31
+ W = float(W)
32
+ k = float(resolution) / min(H, W)
33
+ H *= k
34
+ W *= k
35
+ H = int(np.round(H / 64.0)) * 64
36
+ W = int(np.round(W / 64.0)) * 64
37
+ img = cv2.resize(input_image, (W, H), interpolation=cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA)
38
+ return img
model/video_diffusion/__init__.py ADDED
File without changes
model/video_diffusion/__pycache__/__init__.cpython-39.pyc ADDED
Binary file (168 Bytes). View file
 
model/video_diffusion/models/__init__.py ADDED
File without changes
model/video_diffusion/models/__pycache__/__init__.cpython-39.pyc ADDED
Binary file (175 Bytes). View file
 
model/video_diffusion/models/__pycache__/attention.cpython-39.pyc ADDED
Binary file (9.27 kB). View file
 
model/video_diffusion/models/__pycache__/controlnet3d.cpython-39.pyc ADDED
Binary file (15.7 kB). View file
 
model/video_diffusion/models/__pycache__/resnet.cpython-39.pyc ADDED
Binary file (11.5 kB). View file
 
model/video_diffusion/models/__pycache__/unet_3d_blocks.cpython-39.pyc ADDED
Binary file (10.7 kB). View file
 
model/video_diffusion/models/__pycache__/unet_3d_blocks_control.cpython-39.pyc ADDED
Binary file (3.82 kB). View file
 
model/video_diffusion/models/__pycache__/unet_3d_condition.cpython-39.pyc ADDED
Binary file (14.5 kB). View file
 
model/video_diffusion/models/attention.py ADDED
@@ -0,0 +1,454 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ from dataclasses import dataclass
16
+ from typing import Optional
17
+
18
+ import torch
19
+ from torch import nn
20
+
21
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
22
+ from diffusers.models.modeling_utils import ModelMixin
23
+ from diffusers.models.attention import FeedForward, CrossAttention, AdaLayerNorm
24
+ from diffusers.utils import BaseOutput
25
+ from diffusers.utils.import_utils import is_xformers_available
26
+ from diffusers.models.cross_attention import XFormersCrossAttnProcessor
27
+ from einops import rearrange
28
+
29
+
30
+ @dataclass
31
+ class SpatioTemporalTransformerModelOutput(BaseOutput):
32
+ """torch.FloatTensor of shape [batch x channel x frames x height x width]"""
33
+
34
+ sample: torch.FloatTensor
35
+
36
+
37
+ if is_xformers_available():
38
+ import xformers
39
+ import xformers.ops
40
+ else:
41
+ xformers = None
42
+
43
+
44
+ class SpatioTemporalTransformerModel(ModelMixin, ConfigMixin):
45
+ @register_to_config
46
+ def __init__(
47
+ self,
48
+ num_attention_heads: int = 16,
49
+ attention_head_dim: int = 88,
50
+ in_channels: Optional[int] = None,
51
+ num_layers: int = 1,
52
+ dropout: float = 0.0,
53
+ norm_num_groups: int = 32,
54
+ cross_attention_dim: Optional[int] = None,
55
+ attention_bias: bool = False,
56
+ activation_fn: str = "geglu",
57
+ num_embeds_ada_norm: Optional[int] = None,
58
+ use_linear_projection: bool = False,
59
+ only_cross_attention: bool = False,
60
+ upcast_attention: bool = False,
61
+ **transformer_kwargs,
62
+ ):
63
+ super().__init__()
64
+ self.use_linear_projection = use_linear_projection
65
+ self.num_attention_heads = num_attention_heads
66
+ self.attention_head_dim = attention_head_dim
67
+ inner_dim = num_attention_heads * attention_head_dim
68
+
69
+ # Define input layers
70
+ self.in_channels = in_channels
71
+
72
+ self.norm = torch.nn.GroupNorm(
73
+ num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
74
+ )
75
+ if use_linear_projection:
76
+ self.proj_in = nn.Linear(in_channels, inner_dim)
77
+ else:
78
+ self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
79
+
80
+ # Define transformers blocks
81
+ self.transformer_blocks = nn.ModuleList(
82
+ [
83
+ SpatioTemporalTransformerBlock(
84
+ inner_dim,
85
+ num_attention_heads,
86
+ attention_head_dim,
87
+ dropout=dropout,
88
+ cross_attention_dim=cross_attention_dim,
89
+ activation_fn=activation_fn,
90
+ num_embeds_ada_norm=num_embeds_ada_norm,
91
+ attention_bias=attention_bias,
92
+ only_cross_attention=only_cross_attention,
93
+ upcast_attention=upcast_attention,
94
+ **transformer_kwargs,
95
+ )
96
+ for d in range(num_layers)
97
+ ]
98
+ )
99
+
100
+ # Define output layers
101
+ if use_linear_projection:
102
+ self.proj_out = nn.Linear(in_channels, inner_dim)
103
+ else:
104
+ self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
105
+
106
+ def forward(
107
+ self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True
108
+ ):
109
+ # 1. Input
110
+ clip_length = None
111
+ is_video = hidden_states.ndim == 5
112
+ if is_video:
113
+ clip_length = hidden_states.shape[2]
114
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
115
+ encoder_hidden_states = encoder_hidden_states.repeat_interleave(clip_length, 0)
116
+
117
+ *_, h, w = hidden_states.shape
118
+ residual = hidden_states
119
+
120
+ hidden_states = self.norm(hidden_states)
121
+ if not self.use_linear_projection:
122
+ hidden_states = self.proj_in(hidden_states)
123
+ hidden_states = rearrange(hidden_states, "b c h w -> b (h w) c")
124
+ else:
125
+ hidden_states = rearrange(hidden_states, "b c h w -> b (h w) c")
126
+ hidden_states = self.proj_in(hidden_states)
127
+
128
+ # 2. Blocks
129
+ for block in self.transformer_blocks:
130
+ hidden_states = block(
131
+ hidden_states,
132
+ encoder_hidden_states=encoder_hidden_states,
133
+ timestep=timestep,
134
+ clip_length=clip_length,
135
+ )
136
+
137
+ # 3. Output
138
+ if not self.use_linear_projection:
139
+ hidden_states = rearrange(hidden_states, "b (h w) c -> b c h w", h=h, w=w).contiguous()
140
+ hidden_states = self.proj_out(hidden_states)
141
+ else:
142
+ hidden_states = self.proj_out(hidden_states)
143
+ hidden_states = rearrange(hidden_states, "b (h w) c -> b c h w", h=h, w=w).contiguous()
144
+
145
+ output = hidden_states + residual
146
+ if is_video:
147
+ output = rearrange(output, "(b f) c h w -> b c f h w", f=clip_length)
148
+
149
+ if not return_dict:
150
+ return (output,)
151
+
152
+ return SpatioTemporalTransformerModelOutput(sample=output)
153
+
154
+
155
+ class SpatioTemporalTransformerBlock(nn.Module):
156
+ def __init__(
157
+ self,
158
+ dim: int,
159
+ num_attention_heads: int,
160
+ attention_head_dim: int,
161
+ dropout=0.0,
162
+ cross_attention_dim: Optional[int] = None,
163
+ activation_fn: str = "geglu",
164
+ num_embeds_ada_norm: Optional[int] = None,
165
+ attention_bias: bool = False,
166
+ only_cross_attention: bool = False,
167
+ upcast_attention: bool = False,
168
+ use_sparse_causal_attention: bool = False,
169
+ use_full_sparse_causal_attention: bool = True,
170
+ temporal_attention_position: str = "after_feedforward",
171
+ use_gamma = False,
172
+ ):
173
+ super().__init__()
174
+ self.only_cross_attention = only_cross_attention
175
+ self.use_ada_layer_norm = num_embeds_ada_norm is not None
176
+ self.use_sparse_causal_attention = use_sparse_causal_attention
177
+ self.use_full_sparse_causal_attention = use_full_sparse_causal_attention
178
+ self.use_gamma = use_gamma
179
+
180
+ self.temporal_attention_position = temporal_attention_position
181
+ temporal_attention_positions = ["after_spatial", "after_cross", "after_feedforward"]
182
+ if temporal_attention_position not in temporal_attention_positions:
183
+ raise ValueError(
184
+ f"`temporal_attention_position` must be one of {temporal_attention_positions}"
185
+ )
186
+
187
+ # 1. Spatial-Attn
188
+ if use_sparse_causal_attention:
189
+ spatial_attention = SparseCausalAttention
190
+ elif use_full_sparse_causal_attention:
191
+ spatial_attention = SparseCausalFullAttention
192
+ else:
193
+ spatial_attention = CrossAttention
194
+
195
+ self.attn1 = spatial_attention(
196
+ query_dim=dim,
197
+ heads=num_attention_heads,
198
+ dim_head=attention_head_dim,
199
+ dropout=dropout,
200
+ bias=attention_bias,
201
+ cross_attention_dim=cross_attention_dim if only_cross_attention else None,
202
+ upcast_attention=upcast_attention,
203
+ processor=XFormersCrossAttnProcessor(),
204
+ ) # is a self-attention
205
+ self.norm1 = (
206
+ AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
207
+ )
208
+ if use_gamma:
209
+ self.attn1_gamma = nn.Parameter(torch.ones(dim))
210
+
211
+ # 2. Cross-Attn
212
+ if cross_attention_dim is not None:
213
+ self.attn2 = CrossAttention(
214
+ query_dim=dim,
215
+ cross_attention_dim=cross_attention_dim,
216
+ heads=num_attention_heads,
217
+ dim_head=attention_head_dim,
218
+ dropout=dropout,
219
+ bias=attention_bias,
220
+ upcast_attention=upcast_attention,
221
+ processor=XFormersCrossAttnProcessor(),
222
+ ) # is self-attn if encoder_hidden_states is none
223
+ self.norm2 = (
224
+ AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
225
+ )
226
+ if use_gamma:
227
+ self.attn2_gamma = nn.Parameter(torch.ones(dim))
228
+ else:
229
+ self.attn2 = None
230
+ self.norm2 = None
231
+
232
+ # 3. Temporal-Attn
233
+ self.attn_temporal = CrossAttention(
234
+ query_dim=dim,
235
+ heads=num_attention_heads,
236
+ dim_head=attention_head_dim,
237
+ dropout=dropout,
238
+ bias=attention_bias,
239
+ upcast_attention=upcast_attention,
240
+ processor=XFormersCrossAttnProcessor()
241
+ )
242
+ zero_module(self.attn_temporal) # 默认参数置0
243
+
244
+ self.norm_temporal = (
245
+ AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
246
+ )
247
+
248
+ # 4. Feed-forward
249
+ self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
250
+ self.norm3 = nn.LayerNorm(dim)
251
+ if use_gamma:
252
+ self.ff_gamma = nn.Parameter(torch.ones(dim))
253
+
254
+
255
+ def forward(
256
+ self,
257
+ hidden_states,
258
+ encoder_hidden_states=None,
259
+ timestep=None,
260
+ attention_mask=None,
261
+ clip_length=None,
262
+ ):
263
+ # 1. Self-Attention
264
+ norm_hidden_states = (
265
+ self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
266
+ )
267
+
268
+ kwargs = dict(
269
+ hidden_states=norm_hidden_states,
270
+ attention_mask=attention_mask,
271
+ )
272
+ if self.only_cross_attention:
273
+ kwargs.update(encoder_hidden_states=encoder_hidden_states)
274
+ if self.use_sparse_causal_attention or self.use_full_sparse_causal_attention:
275
+ kwargs.update(clip_length=clip_length)
276
+
277
+ if self.use_gamma:
278
+ hidden_states = hidden_states + self.attn1(**kwargs) * self.attn1_gamma # NOTE gamma
279
+ else:
280
+ hidden_states = hidden_states + self.attn1(**kwargs)
281
+
282
+
283
+ if clip_length is not None and self.temporal_attention_position == "after_spatial":
284
+ hidden_states = self.apply_temporal_attention(hidden_states, timestep, clip_length)
285
+
286
+ if self.attn2 is not None:
287
+ # 2. Cross-Attention
288
+ norm_hidden_states = (
289
+ self.norm2(hidden_states, timestep)
290
+ if self.use_ada_layer_norm
291
+ else self.norm2(hidden_states)
292
+ )
293
+ if self.use_gamma:
294
+ hidden_states = (
295
+ self.attn2(
296
+ norm_hidden_states,
297
+ encoder_hidden_states=encoder_hidden_states,
298
+ attention_mask=attention_mask,
299
+ ) * self.attn2_gamma
300
+ + hidden_states
301
+ )
302
+ else:
303
+ hidden_states = (
304
+ self.attn2(
305
+ norm_hidden_states,
306
+ encoder_hidden_states=encoder_hidden_states,
307
+ attention_mask=attention_mask,
308
+ )
309
+ + hidden_states
310
+ )
311
+
312
+ if clip_length is not None and self.temporal_attention_position == "after_cross":
313
+ hidden_states = self.apply_temporal_attention(hidden_states, timestep, clip_length)
314
+
315
+ # 3. Feed-forward
316
+ if self.use_gamma:
317
+ hidden_states = self.ff(self.norm3(hidden_states)) * self.ff_gamma + hidden_states
318
+ else:
319
+ hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
320
+
321
+ if clip_length is not None and self.temporal_attention_position == "after_feedforward":
322
+ hidden_states = self.apply_temporal_attention(hidden_states, timestep, clip_length)
323
+
324
+ return hidden_states
325
+
326
+ def apply_temporal_attention(self, hidden_states, timestep, clip_length):
327
+ d = hidden_states.shape[1]
328
+ hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=clip_length)
329
+ norm_hidden_states = (
330
+ self.norm_temporal(hidden_states, timestep)
331
+ if self.use_ada_layer_norm
332
+ else self.norm_temporal(hidden_states)
333
+ )
334
+ hidden_states = self.attn_temporal(norm_hidden_states) + hidden_states
335
+ hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
336
+ return hidden_states
337
+
338
+
339
+ class SparseCausalAttention(CrossAttention):
340
+ def forward(
341
+ self,
342
+ hidden_states,
343
+ encoder_hidden_states=None,
344
+ attention_mask=None,
345
+ clip_length: int = None,
346
+ ):
347
+ if (
348
+ self.added_kv_proj_dim is not None
349
+ or encoder_hidden_states is not None
350
+ or attention_mask is not None
351
+ ):
352
+ raise NotImplementedError
353
+
354
+ if self.group_norm is not None:
355
+ hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
356
+
357
+ query = self.to_q(hidden_states)
358
+ dim = query.shape[-1]
359
+ query = self.head_to_batch_dim(query) # 64 4096 40
360
+
361
+ key = self.to_k(hidden_states)
362
+ value = self.to_v(hidden_states)
363
+
364
+ if clip_length is not None and clip_length > 1:
365
+ # spatial temporal
366
+ prev_frame_index = torch.arange(clip_length) - 1
367
+ prev_frame_index[0] = 0
368
+ key = rearrange(key, "(b f) d c -> b f d c", f=clip_length)
369
+ key = torch.cat([key[:, [0] * clip_length], key[:, prev_frame_index]], dim=2)
370
+ key = rearrange(key, "b f d c -> (b f) d c", f=clip_length)
371
+
372
+ value = rearrange(value, "(b f) d c -> b f d c", f=clip_length)
373
+ value = torch.cat([value[:, [0] * clip_length], value[:, prev_frame_index]], dim=2)
374
+ value = rearrange(value, "b f d c -> (b f) d c", f=clip_length)
375
+
376
+
377
+ key = self.head_to_batch_dim(key)
378
+ value = self.head_to_batch_dim(value)
379
+ # use xfromers by default~
380
+ hidden_states = xformers.ops.memory_efficient_attention(
381
+ query, key, value, attn_bias=attention_mask, op=None
382
+ )
383
+ hidden_states = hidden_states.to(query.dtype)
384
+ hidden_states = self.batch_to_head_dim(hidden_states)
385
+
386
+ # linear proj
387
+ hidden_states = self.to_out[0](hidden_states)
388
+
389
+ # dropout
390
+ hidden_states = self.to_out[1](hidden_states)
391
+ return hidden_states
392
+
393
+ def zero_module(module):
394
+ for p in module.parameters():
395
+ nn.init.zeros_(p)
396
+ return module
397
+
398
+
399
+ class SparseCausalFullAttention(CrossAttention):
400
+ def forward(
401
+ self,
402
+ hidden_states,
403
+ encoder_hidden_states=None,
404
+ attention_mask=None,
405
+ clip_length: int = None,
406
+ ):
407
+ if (
408
+ self.added_kv_proj_dim is not None
409
+ or encoder_hidden_states is not None
410
+ or attention_mask is not None
411
+ ):
412
+ raise NotImplementedError
413
+
414
+ if self.group_norm is not None:
415
+ hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
416
+
417
+ query = self.to_q(hidden_states)
418
+ dim = query.shape[-1]
419
+ query = self.head_to_batch_dim(query) # 64 4096 40
420
+
421
+ key = self.to_k(hidden_states)
422
+ value = self.to_v(hidden_states)
423
+
424
+ if clip_length is not None and clip_length > 1:
425
+ # 和所有帧做 spatial temporal attention
426
+ key = rearrange(key, "(b f) d c -> b f d c", f=clip_length)
427
+ # cat full frames
428
+ key = torch.cat([key[:, [iii] * clip_length] for iii in range(clip_length) ], dim=2) # concat第一帧+第i帧。以此为key, value。而非自己这一帧。
429
+ key = rearrange(key, "b f d c -> (b f) d c", f=clip_length)
430
+
431
+ value = rearrange(value, "(b f) d c -> b f d c", f=clip_length)
432
+ value = torch.cat([value[:, [iii] * clip_length] for iii in range(clip_length) ], dim=2) # concat第一帧+第i帧。以此为key, value。而非自己这一帧。
433
+ value = rearrange(value, "b f d c -> (b f) d c", f=clip_length)
434
+
435
+ key = self.head_to_batch_dim(key)
436
+ value = self.head_to_batch_dim(value)
437
+ # use xfromers by default~
438
+ hidden_states = xformers.ops.memory_efficient_attention(
439
+ query, key, value, attn_bias=attention_mask, op=None
440
+ )
441
+ hidden_states = hidden_states.to(query.dtype)
442
+ hidden_states = self.batch_to_head_dim(hidden_states)
443
+
444
+ # linear proj
445
+ hidden_states = self.to_out[0](hidden_states)
446
+
447
+ # dropout
448
+ hidden_states = self.to_out[1](hidden_states)
449
+ return hidden_states
450
+
451
+ def zero_module(module):
452
+ for p in module.parameters():
453
+ nn.init.zeros_(p)
454
+ return module
model/video_diffusion/models/controlnet3d.py ADDED
@@ -0,0 +1,580 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+
16
+ from dataclasses import dataclass
17
+ from typing import Any, Dict, List, Optional, Tuple, Union
18
+
19
+ import torch
20
+ from torch import nn
21
+ from torch.nn import functional as F
22
+
23
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
24
+ from diffusers.utils import BaseOutput, logging
25
+ from diffusers.models.cross_attention import AttnProcessor
26
+ from diffusers.models.embeddings import TimestepEmbedding, Timesteps
27
+ from diffusers.models.modeling_utils import ModelMixin
28
+
29
+ from .unet_3d_blocks import (
30
+ CrossAttnDownBlockPseudo3D,
31
+ DownBlockPseudo3D,
32
+ UNetMidBlockPseudo3DCrossAttn,
33
+ get_down_block,
34
+ )
35
+ from .resnet import PseudoConv3d
36
+ from diffusers.models.cross_attention import AttnProcessor
37
+ from typing import Dict
38
+ from .unet_3d_blocks_control import ControlNetPseudoZeroConv3dBlock, ControlNetInputHintBlock
39
+ import glob
40
+ import os
41
+ import json
42
+
43
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
44
+ @dataclass
45
+ class ControlNetOutput(BaseOutput):
46
+ down_block_res_samples: Tuple[torch.Tensor]
47
+ mid_block_res_sample: torch.Tensor
48
+
49
+
50
+ class ControlNetConditioningEmbedding(nn.Module):
51
+ """
52
+ Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
53
+ [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
54
+ training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
55
+ convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
56
+ (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
57
+ model) to encode image-space conditions ... into feature maps ..."
58
+ """
59
+
60
+ def __init__(
61
+ self,
62
+ conditioning_embedding_channels: int,
63
+ conditioning_channels: int = 3,
64
+ block_out_channels: Tuple[int] = (16, 32, 96, 256),
65
+ ):
66
+ super().__init__()
67
+
68
+ self.conv_in = PseudoConv3d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
69
+
70
+ self.blocks = nn.ModuleList([])
71
+
72
+ for i in range(len(block_out_channels) - 1):
73
+ channel_in = block_out_channels[i]
74
+ channel_out = block_out_channels[i + 1]
75
+ self.blocks.append(PseudoConv3d(channel_in, channel_in, kernel_size=3, padding=1))
76
+ self.blocks.append(PseudoConv3d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
77
+
78
+ # self.conv_out = zero_module(
79
+ # PseudoConv3d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
80
+ # )
81
+ self.conv_out = PseudoConv3d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
82
+
83
+ def forward(self, conditioning):
84
+ embedding = self.conv_in(conditioning)
85
+ embedding = F.silu(embedding)
86
+
87
+ for block in self.blocks:
88
+ embedding = block(embedding)
89
+ embedding = F.silu(embedding)
90
+
91
+ embedding = self.conv_out(embedding)
92
+
93
+ return embedding
94
+
95
+
96
+ class ControlNet3DModel(ModelMixin, ConfigMixin):
97
+ _supports_gradient_checkpointing = True
98
+
99
+ @register_to_config
100
+ def __init__(
101
+ self,
102
+ in_channels: int = 4,
103
+ flip_sin_to_cos: bool = True,
104
+ freq_shift: int = 0,
105
+ down_block_types: Tuple[str] = (
106
+ "CrossAttnDownBlockPseudo3D",
107
+ "CrossAttnDownBlockPseudo3D",
108
+ "CrossAttnDownBlockPseudo3D",
109
+ "DownBlockPseudo3D",
110
+ ),
111
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
112
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
113
+ layers_per_block: int = 2,
114
+ downsample_padding: int = 1,
115
+ mid_block_scale_factor: float = 1,
116
+ act_fn: str = "silu",
117
+ norm_num_groups: Optional[int] = 32,
118
+ norm_eps: float = 1e-5,
119
+ cross_attention_dim: int = 1280,
120
+ attention_head_dim: Union[int, Tuple[int]] = 8,
121
+ use_linear_projection: bool = False,
122
+ class_embed_type: Optional[str] = None,
123
+ num_class_embeds: Optional[int] = None,
124
+ upcast_attention: bool = False,
125
+ resnet_time_scale_shift: str = "default",
126
+ projection_class_embeddings_input_dim: Optional[int] = None,
127
+ controlnet_conditioning_channel_order: str = "rgb",
128
+ conditioning_embedding_out_channels: Optional[Tuple[int]] = (16, 32, 96, 256),
129
+ ):
130
+ super().__init__()
131
+
132
+ # Check inputs
133
+ if len(block_out_channels) != len(down_block_types):
134
+ raise ValueError(
135
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
136
+ )
137
+
138
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
139
+ raise ValueError(
140
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
141
+ )
142
+
143
+ if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
144
+ raise ValueError(
145
+ f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
146
+ )
147
+
148
+ # input
149
+ conv_in_kernel = 3
150
+ conv_in_padding = (conv_in_kernel - 1) // 2
151
+ self.conv_in = PseudoConv3d(
152
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
153
+ )
154
+
155
+ # time
156
+ time_embed_dim = block_out_channels[0] * 4
157
+
158
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
159
+ timestep_input_dim = block_out_channels[0]
160
+
161
+ self.time_embedding = TimestepEmbedding(
162
+ timestep_input_dim,
163
+ time_embed_dim,
164
+ act_fn=act_fn,
165
+ )
166
+
167
+ # class embedding
168
+ if class_embed_type is None and num_class_embeds is not None:
169
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
170
+ elif class_embed_type == "timestep":
171
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
172
+ elif class_embed_type == "identity":
173
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
174
+ elif class_embed_type == "projection":
175
+ if projection_class_embeddings_input_dim is None:
176
+ raise ValueError(
177
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
178
+ )
179
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
180
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
181
+ # 2. it projects from an arbitrary input dimension.
182
+ #
183
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
184
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
185
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
186
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
187
+ else:
188
+ self.class_embedding = None
189
+
190
+ # control net conditioning embedding
191
+ self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
192
+ conditioning_embedding_channels=block_out_channels[0],
193
+ block_out_channels=conditioning_embedding_out_channels,
194
+ )
195
+
196
+ self.down_blocks = nn.ModuleList([])
197
+ self.controlnet_down_blocks = nn.ModuleList([])
198
+
199
+ if isinstance(only_cross_attention, bool):
200
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
201
+
202
+ if isinstance(attention_head_dim, int):
203
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
204
+
205
+ # down
206
+ output_channel = block_out_channels[0]
207
+
208
+ controlnet_block = PseudoConv3d(output_channel, output_channel, kernel_size=1)
209
+ # controlnet_block = zero_module(controlnet_block)
210
+ self.controlnet_down_blocks.append(controlnet_block)
211
+
212
+ for i, down_block_type in enumerate(down_block_types):
213
+ input_channel = output_channel
214
+ output_channel = block_out_channels[i]
215
+ is_final_block = i == len(block_out_channels) - 1
216
+
217
+ down_block = get_down_block(
218
+ down_block_type,
219
+ num_layers=layers_per_block,
220
+ in_channels=input_channel,
221
+ out_channels=output_channel,
222
+ temb_channels=time_embed_dim,
223
+ add_downsample=not is_final_block,
224
+ resnet_eps=norm_eps,
225
+ resnet_act_fn=act_fn,
226
+ resnet_groups=norm_num_groups,
227
+ cross_attention_dim=cross_attention_dim,
228
+ attn_num_head_channels=attention_head_dim[i],
229
+ downsample_padding=downsample_padding,
230
+ use_linear_projection=use_linear_projection,
231
+ only_cross_attention=only_cross_attention[i],
232
+ upcast_attention=upcast_attention,
233
+ resnet_time_scale_shift=resnet_time_scale_shift,
234
+ )
235
+ self.down_blocks.append(down_block)
236
+
237
+ for _ in range(layers_per_block):
238
+ controlnet_block = PseudoConv3d(output_channel, output_channel, kernel_size=1)
239
+ # controlnet_block = zero_module(controlnet_block)
240
+ self.controlnet_down_blocks.append(controlnet_block)
241
+
242
+ if not is_final_block:
243
+ controlnet_block = PseudoConv3d(output_channel, output_channel, kernel_size=1)
244
+ # controlnet_block = zero_module(controlnet_block)
245
+ self.controlnet_down_blocks.append(controlnet_block)
246
+
247
+ # mid
248
+ mid_block_channel = block_out_channels[-1]
249
+
250
+ controlnet_block = PseudoConv3d(mid_block_channel, mid_block_channel, kernel_size=1)
251
+ # controlnet_block = zero_module(controlnet_block)
252
+ self.controlnet_mid_block = controlnet_block
253
+
254
+ self.mid_block = UNetMidBlockPseudo3DCrossAttn(
255
+ in_channels=mid_block_channel,
256
+ temb_channels=time_embed_dim,
257
+ resnet_eps=norm_eps,
258
+ resnet_act_fn=act_fn,
259
+ output_scale_factor=mid_block_scale_factor,
260
+ resnet_time_scale_shift=resnet_time_scale_shift,
261
+ cross_attention_dim=cross_attention_dim,
262
+ attn_num_head_channels=attention_head_dim[-1],
263
+ resnet_groups=norm_num_groups,
264
+ use_linear_projection=use_linear_projection,
265
+ upcast_attention=upcast_attention,
266
+ )
267
+
268
+ @property
269
+ # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
270
+ def attn_processors(self) -> Dict[str, AttnProcessor]:
271
+ r"""
272
+ Returns:
273
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
274
+ indexed by its weight name.
275
+ """
276
+ # set recursively
277
+ processors = {}
278
+
279
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttnProcessor]):
280
+ if hasattr(module, "set_processor"):
281
+ processors[f"{name}.processor"] = module.processor
282
+
283
+ for sub_name, child in module.named_children():
284
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
285
+
286
+ return processors
287
+
288
+ for name, module in self.named_children():
289
+ fn_recursive_add_processors(name, module, processors)
290
+
291
+ return processors
292
+
293
+ # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
294
+ def set_attn_processor(self, processor: Union[AttnProcessor, Dict[str, AttnProcessor]]):
295
+ r"""
296
+ Parameters:
297
+ `processor (`dict` of `AttnProcessor` or `AttnProcessor`):
298
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
299
+ of **all** `CrossAttention` layers.
300
+ In case `processor` is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainablae attention processors.:
301
+
302
+ """
303
+ count = len(self.attn_processors.keys())
304
+
305
+ if isinstance(processor, dict) and len(processor) != count:
306
+ raise ValueError(
307
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
308
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
309
+ )
310
+
311
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
312
+ if hasattr(module, "set_processor"):
313
+ if not isinstance(processor, dict):
314
+ module.set_processor(processor)
315
+ else:
316
+ module.set_processor(processor.pop(f"{name}.processor"))
317
+
318
+ for sub_name, child in module.named_children():
319
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
320
+
321
+ for name, module in self.named_children():
322
+ fn_recursive_attn_processor(name, module, processor)
323
+
324
+ # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
325
+ def set_attention_slice(self, slice_size):
326
+ r"""
327
+ Enable sliced attention computation.
328
+
329
+ When this option is enabled, the attention module will split the input tensor in slices, to compute attention
330
+ in several steps. This is useful to save some memory in exchange for a small speed decrease.
331
+
332
+ Args:
333
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
334
+ When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
335
+ `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
336
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
337
+ must be a multiple of `slice_size`.
338
+ """
339
+ sliceable_head_dims = []
340
+
341
+ def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
342
+ if hasattr(module, "set_attention_slice"):
343
+ sliceable_head_dims.append(module.sliceable_head_dim)
344
+
345
+ for child in module.children():
346
+ fn_recursive_retrieve_slicable_dims(child)
347
+
348
+ # retrieve number of attention layers
349
+ for module in self.children():
350
+ fn_recursive_retrieve_slicable_dims(module)
351
+
352
+ num_slicable_layers = len(sliceable_head_dims)
353
+
354
+ if slice_size == "auto":
355
+ # half the attention head size is usually a good trade-off between
356
+ # speed and memory
357
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
358
+ elif slice_size == "max":
359
+ # make smallest slice possible
360
+ slice_size = num_slicable_layers * [1]
361
+
362
+ slice_size = num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
363
+
364
+ if len(slice_size) != len(sliceable_head_dims):
365
+ raise ValueError(
366
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
367
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
368
+ )
369
+
370
+ for i in range(len(slice_size)):
371
+ size = slice_size[i]
372
+ dim = sliceable_head_dims[i]
373
+ if size is not None and size > dim:
374
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
375
+
376
+ # Recursively walk through all the children.
377
+ # Any children which exposes the set_attention_slice method
378
+ # gets the message
379
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
380
+ if hasattr(module, "set_attention_slice"):
381
+ module.set_attention_slice(slice_size.pop())
382
+
383
+ for child in module.children():
384
+ fn_recursive_set_attention_slice(child, slice_size)
385
+
386
+ reversed_slice_size = list(reversed(slice_size))
387
+ for module in self.children():
388
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
389
+
390
+ def _set_gradient_checkpointing(self, module, value=False):
391
+ if isinstance(module, (CrossAttnDownBlockPseudo3D, DownBlockPseudo3D)):
392
+ module.gradient_checkpointing = value
393
+
394
+ def forward(
395
+ self,
396
+ sample: torch.FloatTensor,
397
+ timestep: Union[torch.Tensor, float, int],
398
+ encoder_hidden_states: torch.Tensor,
399
+ controlnet_cond: torch.FloatTensor,
400
+ class_labels: Optional[torch.Tensor] = None,
401
+ timestep_cond: Optional[torch.Tensor] = None,
402
+ attention_mask: Optional[torch.Tensor] = None,
403
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
404
+ return_dict: bool = True,
405
+ ) -> Union[ControlNetOutput, Tuple]:
406
+ # check channel order
407
+ channel_order = self.config.controlnet_conditioning_channel_order
408
+
409
+ if channel_order == "rgb":
410
+ # in rgb order by default
411
+ ...
412
+ elif channel_order == "bgr":
413
+ controlnet_cond = torch.flip(controlnet_cond, dims=[1])
414
+ else:
415
+ raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
416
+
417
+ # prepare attention_mask
418
+ if attention_mask is not None:
419
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
420
+ attention_mask = attention_mask.unsqueeze(1)
421
+
422
+ # 1. time
423
+ timesteps = timestep
424
+ if not torch.is_tensor(timesteps):
425
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
426
+ # This would be a good case for the `match` statement (Python 3.10+)
427
+ is_mps = sample.device.type == "mps"
428
+ if isinstance(timestep, float):
429
+ dtype = torch.float32 if is_mps else torch.float64
430
+ else:
431
+ dtype = torch.int32 if is_mps else torch.int64
432
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
433
+ elif len(timesteps.shape) == 0:
434
+ timesteps = timesteps[None].to(sample.device)
435
+
436
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
437
+ timesteps = timesteps.expand(sample.shape[0])
438
+
439
+ t_emb = self.time_proj(timesteps)
440
+
441
+ # timesteps does not contain any weights and will always return f32 tensors
442
+ # but time_embedding might actually be running in fp16. so we need to cast here.
443
+ # there might be better ways to encapsulate this.
444
+ t_emb = t_emb.to(dtype=self.dtype)
445
+
446
+ emb = self.time_embedding(t_emb, timestep_cond)
447
+
448
+ if self.class_embedding is not None:
449
+ if class_labels is None:
450
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
451
+
452
+ if self.config.class_embed_type == "timestep":
453
+ class_labels = self.time_proj(class_labels)
454
+
455
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
456
+ emb = emb + class_emb
457
+
458
+ # 2. pre-process
459
+ sample = self.conv_in(sample)
460
+
461
+ controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
462
+ # print(sample.shape, controlnet_cond.shape)
463
+
464
+ sample += controlnet_cond
465
+ # 3. down
466
+
467
+ down_block_res_samples = (sample,)
468
+ for downsample_block in self.down_blocks:
469
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
470
+ sample, res_samples = downsample_block(
471
+ hidden_states=sample,
472
+ temb=emb,
473
+ encoder_hidden_states=encoder_hidden_states,
474
+ attention_mask=attention_mask,
475
+ )
476
+ else:
477
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
478
+
479
+ down_block_res_samples += res_samples
480
+
481
+ # 4. mid
482
+ if self.mid_block is not None:
483
+ sample = self.mid_block(
484
+ sample,
485
+ emb,
486
+ encoder_hidden_states=encoder_hidden_states,
487
+ attention_mask=attention_mask,
488
+ )
489
+
490
+ # 5. Control net blocks
491
+
492
+ controlnet_down_block_res_samples = ()
493
+
494
+ for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
495
+ down_block_res_sample = controlnet_block(down_block_res_sample)
496
+ controlnet_down_block_res_samples += (down_block_res_sample,)
497
+
498
+ down_block_res_samples = controlnet_down_block_res_samples
499
+
500
+ mid_block_res_sample = self.controlnet_mid_block(sample)
501
+
502
+ if not return_dict:
503
+ return (down_block_res_samples, mid_block_res_sample)
504
+
505
+ return ControlNetOutput(
506
+ down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
507
+ )
508
+
509
+ @classmethod
510
+ def from_2d_model(cls, model_path, condition_on_fps=False, controlnet_hint_channels: Optional[int] = None,):
511
+ '''
512
+ load a 2d model and convert it to a pseudo 3d model
513
+ '''
514
+ config_path = os.path.join(model_path, "config.json")
515
+ if not os.path.isfile(config_path):
516
+ raise RuntimeError(f"{config_path} does not exist")
517
+ with open(config_path, "r") as f:
518
+ config = json.load(f)
519
+
520
+ config.pop("_class_name")
521
+ config.pop("_diffusers_version")
522
+
523
+ block_replacer = {
524
+ "CrossAttnDownBlock2D": "CrossAttnDownBlockPseudo3D",
525
+ "DownBlock2D": "DownBlockPseudo3D",
526
+ "UNetMidBlock2DCrossAttn": "UNetMidBlockPseudo3DCrossAttn",
527
+ }
528
+
529
+ def convert_2d_to_3d_block(block):
530
+ return block_replacer[block] if block in block_replacer else block
531
+
532
+ config["down_block_types"] = [
533
+ convert_2d_to_3d_block(block) for block in config["down_block_types"]
534
+ ]
535
+
536
+ if "mid_block_type" in config:
537
+ config["mid_block_type"] = convert_2d_to_3d_block(config["mid_block_type"])
538
+
539
+ if condition_on_fps:
540
+ config["fps_embed_type"] = "timestep" # 和timestep保持一致的type。
541
+
542
+ if controlnet_hint_channels:
543
+ config["controlnet_hint_channels"] = controlnet_hint_channels
544
+
545
+ print(config)
546
+
547
+ model = cls(**config) # 调用自身(init), 传入config参数全换成3d的setting
548
+ state_dict_path_condidates = glob.glob(os.path.join(model_path, "*.bin"))
549
+ if state_dict_path_condidates:
550
+ state_dict = torch.load(state_dict_path_condidates[0], map_location="cpu")
551
+ model.load_2d_state_dict(state_dict=state_dict)
552
+
553
+ return model
554
+
555
+ def load_2d_state_dict(self, state_dict, **kwargs):
556
+ '''
557
+ 2D 部分的参数名完全不变。
558
+ '''
559
+ state_dict_3d = self.state_dict()
560
+ # print("diff params list:", list(set(state_dict_3d.keys()) - set(state_dict.keys())))
561
+
562
+ for k, v in state_dict.items():
563
+ if k not in state_dict_3d:
564
+ raise KeyError(f"2d state_dict key {k} does not exist in 3d model")
565
+
566
+ for k, v in state_dict_3d.items():
567
+ if "_temporal" in k:
568
+ continue
569
+ if "gamma" in k:
570
+ continue
571
+ if k not in state_dict:
572
+ raise KeyError(f"3d state_dict key {k} does not exist in 2d model")
573
+ state_dict_3d.update(state_dict)
574
+ self.load_state_dict(state_dict_3d, strict=True, **kwargs)
575
+
576
+
577
+ def zero_module(module):
578
+ for p in module.parameters():
579
+ nn.init.zeros_(p)
580
+ return module
model/video_diffusion/models/resnet.py ADDED
@@ -0,0 +1,486 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ from functools import partial
16
+
17
+ import torch
18
+ import torch.nn as nn
19
+ import torch.nn.functional as F
20
+
21
+ from einops import rearrange
22
+
23
+
24
+ class PseudoConv3d(nn.Conv2d):
25
+ def __init__(self, in_channels, out_channels, kernel_size, temporal_kernel_size=None, **kwargs):
26
+ super().__init__(
27
+ in_channels=in_channels,
28
+ out_channels=out_channels,
29
+ kernel_size=kernel_size,
30
+ **kwargs,
31
+ )
32
+ if temporal_kernel_size is None:
33
+ temporal_kernel_size = kernel_size
34
+
35
+ self.conv_temporal = (
36
+ nn.Conv1d(
37
+ out_channels,
38
+ out_channels,
39
+ kernel_size=temporal_kernel_size,
40
+ padding=temporal_kernel_size // 2,
41
+ )
42
+ if kernel_size > 1
43
+ else None
44
+ )
45
+
46
+ if self.conv_temporal is not None:
47
+ nn.init.dirac_(self.conv_temporal.weight.data) # initialized to be identity
48
+ nn.init.zeros_(self.conv_temporal.bias.data)
49
+
50
+ def forward(self, x):
51
+ b = x.shape[0]
52
+
53
+ is_video = x.ndim == 5
54
+ if is_video:
55
+ x = rearrange(x, "b c f h w -> (b f) c h w")
56
+
57
+ x = super().forward(x)
58
+
59
+ if is_video:
60
+ x = rearrange(x, "(b f) c h w -> b c f h w", b=b)
61
+
62
+ if self.conv_temporal is None or not is_video:
63
+ return x
64
+
65
+ *_, h, w = x.shape
66
+
67
+ x = rearrange(x, "b c f h w -> (b h w) c f")
68
+
69
+ x = self.conv_temporal(x) # 加入空间1D的时序卷积。channel不变。(建模时序信息)
70
+
71
+ x = rearrange(x, "(b h w) c f -> b c f h w", h=h, w=w)
72
+
73
+ return x
74
+
75
+
76
+ class UpsamplePseudo3D(nn.Module):
77
+ """
78
+ An upsampling layer with an optional convolution.
79
+
80
+ Parameters:
81
+ channels: channels in the inputs and outputs.
82
+ use_conv: a bool determining if a convolution is applied.
83
+ use_conv_transpose:
84
+ out_channels:
85
+ """
86
+
87
+ def __init__(
88
+ self, channels, use_conv=False, use_conv_transpose=False, out_channels=None, name="conv"
89
+ ):
90
+ super().__init__()
91
+ self.channels = channels
92
+ self.out_channels = out_channels or channels
93
+ self.use_conv = use_conv
94
+ self.use_conv_transpose = use_conv_transpose
95
+ self.name = name
96
+
97
+ conv = None
98
+ if use_conv_transpose:
99
+ raise NotImplementedError
100
+ conv = nn.ConvTranspose2d(channels, self.out_channels, 4, 2, 1)
101
+ elif use_conv:
102
+ conv = PseudoConv3d(self.channels, self.out_channels, 3, padding=1)
103
+
104
+ # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
105
+ if name == "conv":
106
+ self.conv = conv
107
+ else:
108
+ self.Conv2d_0 = conv
109
+
110
+ def forward(self, hidden_states, output_size=None):
111
+ assert hidden_states.shape[1] == self.channels
112
+
113
+ # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
114
+ # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
115
+ # https://github.com/pytorch/pytorch/issues/86679
116
+ dtype = hidden_states.dtype
117
+ if dtype == torch.bfloat16:
118
+ hidden_states = hidden_states.to(torch.float32)
119
+
120
+ # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
121
+ if hidden_states.shape[0] >= 64:
122
+ hidden_states = hidden_states.contiguous()
123
+
124
+ b = hidden_states.shape[0]
125
+ is_video = hidden_states.ndim == 5
126
+ if is_video:
127
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
128
+
129
+ # if `output_size` is passed we force the interpolation output
130
+ # size and do not make use of `scale_factor=2`
131
+ if output_size is None:
132
+ # 先插值再用conv
133
+ hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
134
+ else:
135
+ hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
136
+
137
+ # If the input is bfloat16, we cast back to bfloat16
138
+ if dtype == torch.bfloat16:
139
+ hidden_states = hidden_states.to(dtype)
140
+
141
+ if is_video:
142
+ hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", b=b)
143
+
144
+ if self.use_conv:
145
+ if self.name == "conv":
146
+ hidden_states = self.conv(hidden_states)
147
+ else:
148
+ hidden_states = self.Conv2d_0(hidden_states)
149
+
150
+ return hidden_states
151
+
152
+
153
+ class DownsamplePseudo3D(nn.Module):
154
+ """
155
+ A downsampling layer with an optional convolution.
156
+
157
+ Parameters:
158
+ channels: channels in the inputs and outputs.
159
+ use_conv: a bool determining if a convolution is applied.
160
+ out_channels:
161
+ padding:
162
+ """
163
+
164
+ def __init__(self, channels, use_conv=False, out_channels=None, padding=1, name="conv"):
165
+ super().__init__()
166
+ self.channels = channels
167
+ self.out_channels = out_channels or channels
168
+ self.use_conv = use_conv
169
+ self.padding = padding
170
+ stride = 2
171
+ self.name = name
172
+
173
+ if use_conv:
174
+ conv = PseudoConv3d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
175
+ else:
176
+ assert self.channels == self.out_channels
177
+ conv = nn.AvgPool2d(kernel_size=stride, stride=stride)
178
+
179
+ if name == "conv":
180
+ self.Conv2d_0 = conv
181
+ self.conv = conv
182
+ elif name == "Conv2d_0":
183
+ self.conv = conv
184
+ else:
185
+ self.conv = conv
186
+
187
+ def forward(self, hidden_states):
188
+ assert hidden_states.shape[1] == self.channels
189
+ if self.use_conv and self.padding == 0:
190
+ pad = (0, 1, 0, 1)
191
+ hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
192
+
193
+ assert hidden_states.shape[1] == self.channels
194
+ if self.use_conv:
195
+ hidden_states = self.conv(hidden_states)
196
+ else:
197
+ b = hidden_states.shape[0]
198
+ is_video = hidden_states.ndim == 5
199
+ if is_video:
200
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
201
+ hidden_states = self.conv(hidden_states)
202
+ if is_video:
203
+ hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", b=b)
204
+
205
+ return hidden_states
206
+
207
+
208
+ class ResnetBlockPseudo3D(nn.Module):
209
+ def __init__(
210
+ self,
211
+ *,
212
+ in_channels,
213
+ out_channels=None,
214
+ conv_shortcut=False,
215
+ dropout=0.0,
216
+ temb_channels=512,
217
+ groups=32,
218
+ groups_out=None,
219
+ pre_norm=True,
220
+ eps=1e-6,
221
+ non_linearity="swish",
222
+ time_embedding_norm="default",
223
+ kernel=None,
224
+ output_scale_factor=1.0,
225
+ use_in_shortcut=None,
226
+ up=False,
227
+ down=False,
228
+ ):
229
+ super().__init__()
230
+ self.pre_norm = pre_norm
231
+ self.pre_norm = True
232
+ self.in_channels = in_channels
233
+ out_channels = in_channels if out_channels is None else out_channels
234
+ self.out_channels = out_channels
235
+ self.use_conv_shortcut = conv_shortcut
236
+ self.time_embedding_norm = time_embedding_norm
237
+ self.up = up
238
+ self.down = down
239
+ self.output_scale_factor = output_scale_factor
240
+
241
+ if groups_out is None:
242
+ groups_out = groups
243
+
244
+ self.norm1 = torch.nn.GroupNorm(
245
+ num_groups=groups, num_channels=in_channels, eps=eps, affine=True
246
+ )
247
+
248
+ self.conv1 = PseudoConv3d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
249
+
250
+ if temb_channels is not None:
251
+ if self.time_embedding_norm == "default":
252
+ time_emb_proj_out_channels = out_channels
253
+ elif self.time_embedding_norm == "scale_shift":
254
+ time_emb_proj_out_channels = out_channels * 2
255
+ else:
256
+ raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
257
+
258
+ self.time_emb_proj = torch.nn.Linear(temb_channels, time_emb_proj_out_channels)
259
+ else:
260
+ self.time_emb_proj = None
261
+
262
+ self.norm2 = torch.nn.GroupNorm(
263
+ num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
264
+ )
265
+ self.dropout = torch.nn.Dropout(dropout)
266
+ self.conv2 = PseudoConv3d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
267
+
268
+ if non_linearity == "swish":
269
+ self.nonlinearity = lambda x: F.silu(x)
270
+ elif non_linearity == "mish":
271
+ self.nonlinearity = Mish()
272
+ elif non_linearity == "silu":
273
+ self.nonlinearity = nn.SiLU()
274
+
275
+ self.upsample = self.downsample = None
276
+ if self.up:
277
+ if kernel == "fir":
278
+ fir_kernel = (1, 3, 3, 1)
279
+ self.upsample = lambda x: upsample_2d(x, kernel=fir_kernel)
280
+ elif kernel == "sde_vp":
281
+ self.upsample = partial(F.interpolate, scale_factor=2.0, mode="nearest")
282
+ else:
283
+ self.upsample = UpsamplePseudo3D(in_channels, use_conv=False)
284
+ elif self.down:
285
+ if kernel == "fir":
286
+ fir_kernel = (1, 3, 3, 1)
287
+ self.downsample = lambda x: downsample_2d(x, kernel=fir_kernel)
288
+ elif kernel == "sde_vp":
289
+ self.downsample = partial(F.avg_pool2d, kernel_size=2, stride=2)
290
+ else:
291
+ self.downsample = DownsamplePseudo3D(in_channels, use_conv=False, padding=1, name="op")
292
+
293
+ self.use_in_shortcut = (
294
+ self.in_channels != self.out_channels if use_in_shortcut is None else use_in_shortcut
295
+ )
296
+
297
+ self.conv_shortcut = None
298
+ if self.use_in_shortcut:
299
+ self.conv_shortcut = PseudoConv3d(
300
+ in_channels, out_channels, kernel_size=1, stride=1, padding=0
301
+ )
302
+
303
+ def forward(self, input_tensor, temb):
304
+ hidden_states = input_tensor
305
+
306
+ hidden_states = self.norm1(hidden_states)
307
+ hidden_states = self.nonlinearity(hidden_states)
308
+
309
+ if self.upsample is not None:
310
+ # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
311
+ if hidden_states.shape[0] >= 64:
312
+ input_tensor = input_tensor.contiguous()
313
+ hidden_states = hidden_states.contiguous()
314
+ input_tensor = self.upsample(input_tensor)
315
+ hidden_states = self.upsample(hidden_states)
316
+ elif self.downsample is not None:
317
+ input_tensor = self.downsample(input_tensor)
318
+ hidden_states = self.downsample(hidden_states)
319
+
320
+ hidden_states = self.conv1(hidden_states)
321
+
322
+ if temb is not None:
323
+ temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
324
+
325
+ if temb is not None and self.time_embedding_norm == "default":
326
+ is_video = hidden_states.ndim == 5
327
+ if is_video:
328
+ b, c, f, h, w = hidden_states.shape
329
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
330
+ temb = temb.repeat_interleave(f, 0)
331
+
332
+ hidden_states = hidden_states + temb
333
+
334
+ if is_video:
335
+ hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", b=b)
336
+
337
+ hidden_states = self.norm2(hidden_states)
338
+
339
+ if temb is not None and self.time_embedding_norm == "scale_shift":
340
+ is_video = hidden_states.ndim == 5
341
+ if is_video:
342
+ b, c, f, h, w = hidden_states.shape
343
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
344
+ temb = temb.repeat_interleave(f, 0)
345
+
346
+ scale, shift = torch.chunk(temb, 2, dim=1)
347
+ hidden_states = hidden_states * (1 + scale) + shift
348
+
349
+ if is_video:
350
+ hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", b=b)
351
+
352
+ hidden_states = self.nonlinearity(hidden_states)
353
+
354
+ hidden_states = self.dropout(hidden_states)
355
+ hidden_states = self.conv2(hidden_states)
356
+
357
+ if self.conv_shortcut is not None:
358
+ input_tensor = self.conv_shortcut(input_tensor)
359
+
360
+ output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
361
+
362
+ return output_tensor
363
+
364
+
365
+ class Mish(torch.nn.Module):
366
+ def forward(self, hidden_states):
367
+ return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))
368
+
369
+
370
+ def upsample_2d(hidden_states, kernel=None, factor=2, gain=1):
371
+ r"""Upsample2D a batch of 2D images with the given filter.
372
+ Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and upsamples each image with the given
373
+ filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified
374
+ `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is
375
+ a: multiple of the upsampling factor.
376
+
377
+ Args:
378
+ hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
379
+ kernel: FIR filter of the shape `[firH, firW]` or `[firN]`
380
+ (separable). The default is `[1] * factor`, which corresponds to nearest-neighbor upsampling.
381
+ factor: Integer upsampling factor (default: 2).
382
+ gain: Scaling factor for signal magnitude (default: 1.0).
383
+
384
+ Returns:
385
+ output: Tensor of the shape `[N, C, H * factor, W * factor]`
386
+ """
387
+ assert isinstance(factor, int) and factor >= 1
388
+ if kernel is None:
389
+ kernel = [1] * factor
390
+
391
+ kernel = torch.tensor(kernel, dtype=torch.float32)
392
+ if kernel.ndim == 1:
393
+ kernel = torch.outer(kernel, kernel)
394
+ kernel /= torch.sum(kernel)
395
+
396
+ kernel = kernel * (gain * (factor**2))
397
+ pad_value = kernel.shape[0] - factor
398
+ output = upfirdn2d_native(
399
+ hidden_states,
400
+ kernel.to(device=hidden_states.device),
401
+ up=factor,
402
+ pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
403
+ )
404
+ return output
405
+
406
+
407
+ def downsample_2d(hidden_states, kernel=None, factor=2, gain=1):
408
+ r"""Downsample2D a batch of 2D images with the given filter.
409
+ Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
410
+ given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
411
+ specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
412
+ shape is a multiple of the downsampling factor.
413
+
414
+ Args:
415
+ hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
416
+ kernel: FIR filter of the shape `[firH, firW]` or `[firN]`
417
+ (separable). The default is `[1] * factor`, which corresponds to average pooling.
418
+ factor: Integer downsampling factor (default: 2).
419
+ gain: Scaling factor for signal magnitude (default: 1.0).
420
+
421
+ Returns:
422
+ output: Tensor of the shape `[N, C, H // factor, W // factor]`
423
+ """
424
+
425
+ assert isinstance(factor, int) and factor >= 1
426
+ if kernel is None:
427
+ kernel = [1] * factor
428
+
429
+ kernel = torch.tensor(kernel, dtype=torch.float32)
430
+ if kernel.ndim == 1:
431
+ kernel = torch.outer(kernel, kernel)
432
+ kernel /= torch.sum(kernel)
433
+
434
+ kernel = kernel * gain
435
+ pad_value = kernel.shape[0] - factor
436
+ output = upfirdn2d_native(
437
+ hidden_states,
438
+ kernel.to(device=hidden_states.device),
439
+ down=factor,
440
+ pad=((pad_value + 1) // 2, pad_value // 2),
441
+ )
442
+ return output
443
+
444
+
445
+ def upfirdn2d_native(tensor, kernel, up=1, down=1, pad=(0, 0)):
446
+ up_x = up_y = up
447
+ down_x = down_y = down
448
+ pad_x0 = pad_y0 = pad[0]
449
+ pad_x1 = pad_y1 = pad[1]
450
+
451
+ _, channel, in_h, in_w = tensor.shape
452
+ tensor = tensor.reshape(-1, in_h, in_w, 1)
453
+
454
+ _, in_h, in_w, minor = tensor.shape
455
+ kernel_h, kernel_w = kernel.shape
456
+
457
+ out = tensor.view(-1, in_h, 1, in_w, 1, minor)
458
+ out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
459
+ out = out.view(-1, in_h * up_y, in_w * up_x, minor)
460
+
461
+ out = F.pad(out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
462
+ out = out.to(tensor.device) # Move back to mps if necessary
463
+ out = out[
464
+ :,
465
+ max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
466
+ max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
467
+ :,
468
+ ]
469
+
470
+ out = out.permute(0, 3, 1, 2)
471
+ out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
472
+ w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
473
+ out = F.conv2d(out, w)
474
+ out = out.reshape(
475
+ -1,
476
+ minor,
477
+ in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
478
+ in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
479
+ )
480
+ out = out.permute(0, 2, 3, 1)
481
+ out = out[:, ::down_y, ::down_x, :]
482
+
483
+ out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
484
+ out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
485
+
486
+ return out.view(-1, channel, out_h, out_w)
model/video_diffusion/models/unet_3d_blocks.py ADDED
@@ -0,0 +1,622 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import torch
16
+ from torch import nn
17
+
18
+ from .attention import SpatioTemporalTransformerModel
19
+ from .resnet import DownsamplePseudo3D, ResnetBlockPseudo3D, UpsamplePseudo3D
20
+
21
+
22
+ def get_down_block(
23
+ down_block_type,
24
+ num_layers,
25
+ in_channels,
26
+ out_channels,
27
+ temb_channels,
28
+ add_downsample,
29
+ resnet_eps,
30
+ resnet_act_fn,
31
+ attn_num_head_channels,
32
+ resnet_groups=None,
33
+ cross_attention_dim=None,
34
+ downsample_padding=None,
35
+ dual_cross_attention=False,
36
+ use_linear_projection=False,
37
+ only_cross_attention=False,
38
+ upcast_attention=False,
39
+ resnet_time_scale_shift="default",
40
+ ):
41
+ down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
42
+ if down_block_type == "DownBlockPseudo3D":
43
+ return DownBlockPseudo3D(
44
+ num_layers=num_layers,
45
+ in_channels=in_channels,
46
+ out_channels=out_channels,
47
+ temb_channels=temb_channels,
48
+ add_downsample=add_downsample,
49
+ resnet_eps=resnet_eps,
50
+ resnet_act_fn=resnet_act_fn,
51
+ resnet_groups=resnet_groups,
52
+ downsample_padding=downsample_padding,
53
+ resnet_time_scale_shift=resnet_time_scale_shift,
54
+ )
55
+ elif down_block_type == "CrossAttnDownBlockPseudo3D":
56
+ if cross_attention_dim is None:
57
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockPseudo3D")
58
+ return CrossAttnDownBlockPseudo3D(
59
+ num_layers=num_layers,
60
+ in_channels=in_channels,
61
+ out_channels=out_channels,
62
+ temb_channels=temb_channels,
63
+ add_downsample=add_downsample,
64
+ resnet_eps=resnet_eps,
65
+ resnet_act_fn=resnet_act_fn,
66
+ resnet_groups=resnet_groups,
67
+ downsample_padding=downsample_padding,
68
+ cross_attention_dim=cross_attention_dim,
69
+ attn_num_head_channels=attn_num_head_channels,
70
+ dual_cross_attention=dual_cross_attention,
71
+ use_linear_projection=use_linear_projection,
72
+ only_cross_attention=only_cross_attention,
73
+ upcast_attention=upcast_attention,
74
+ resnet_time_scale_shift=resnet_time_scale_shift,
75
+ )
76
+ raise ValueError(f"{down_block_type} does not exist.")
77
+
78
+
79
+ def get_up_block(
80
+ up_block_type,
81
+ num_layers,
82
+ in_channels,
83
+ out_channels,
84
+ prev_output_channel,
85
+ temb_channels,
86
+ add_upsample,
87
+ resnet_eps,
88
+ resnet_act_fn,
89
+ attn_num_head_channels,
90
+ resnet_groups=None,
91
+ cross_attention_dim=None,
92
+ dual_cross_attention=False,
93
+ use_linear_projection=False,
94
+ only_cross_attention=False,
95
+ upcast_attention=False,
96
+ resnet_time_scale_shift="default",
97
+ ):
98
+ up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
99
+ if up_block_type == "UpBlockPseudo3D":
100
+ return UpBlockPseudo3D(
101
+ num_layers=num_layers,
102
+ in_channels=in_channels,
103
+ out_channels=out_channels,
104
+ prev_output_channel=prev_output_channel,
105
+ temb_channels=temb_channels,
106
+ add_upsample=add_upsample,
107
+ resnet_eps=resnet_eps,
108
+ resnet_act_fn=resnet_act_fn,
109
+ resnet_groups=resnet_groups,
110
+ resnet_time_scale_shift=resnet_time_scale_shift,
111
+ )
112
+ elif up_block_type == "CrossAttnUpBlockPseudo3D":
113
+ if cross_attention_dim is None:
114
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockPseudo3D")
115
+ return CrossAttnUpBlockPseudo3D(
116
+ num_layers=num_layers,
117
+ in_channels=in_channels,
118
+ out_channels=out_channels,
119
+ prev_output_channel=prev_output_channel,
120
+ temb_channels=temb_channels,
121
+ add_upsample=add_upsample,
122
+ resnet_eps=resnet_eps,
123
+ resnet_act_fn=resnet_act_fn,
124
+ resnet_groups=resnet_groups,
125
+ cross_attention_dim=cross_attention_dim,
126
+ attn_num_head_channels=attn_num_head_channels,
127
+ dual_cross_attention=dual_cross_attention,
128
+ use_linear_projection=use_linear_projection,
129
+ only_cross_attention=only_cross_attention,
130
+ upcast_attention=upcast_attention,
131
+ resnet_time_scale_shift=resnet_time_scale_shift,
132
+ )
133
+ raise ValueError(f"{up_block_type} does not exist.")
134
+
135
+
136
+ class UNetMidBlockPseudo3DCrossAttn(nn.Module):
137
+ def __init__(
138
+ self,
139
+ in_channels: int,
140
+ temb_channels: int,
141
+ dropout: float = 0.0,
142
+ num_layers: int = 1,
143
+ resnet_eps: float = 1e-6,
144
+ resnet_time_scale_shift: str = "default",
145
+ resnet_act_fn: str = "swish",
146
+ resnet_groups: int = 32,
147
+ resnet_pre_norm: bool = True,
148
+ attn_num_head_channels=1,
149
+ output_scale_factor=1.0,
150
+ cross_attention_dim=1280,
151
+ dual_cross_attention=False,
152
+ use_linear_projection=False,
153
+ upcast_attention=False,
154
+ ):
155
+ super().__init__()
156
+
157
+ self.has_cross_attention = True
158
+ self.attn_num_head_channels = attn_num_head_channels
159
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
160
+
161
+ # there is always at least one resnet
162
+ resnets = [
163
+ ResnetBlockPseudo3D(
164
+ in_channels=in_channels,
165
+ out_channels=in_channels,
166
+ temb_channels=temb_channels,
167
+ eps=resnet_eps,
168
+ groups=resnet_groups,
169
+ dropout=dropout,
170
+ time_embedding_norm=resnet_time_scale_shift,
171
+ non_linearity=resnet_act_fn,
172
+ output_scale_factor=output_scale_factor,
173
+ pre_norm=resnet_pre_norm,
174
+ )
175
+ ]
176
+ attentions = []
177
+
178
+ for _ in range(num_layers):
179
+ if dual_cross_attention:
180
+ raise NotImplementedError
181
+ attentions.append(
182
+ SpatioTemporalTransformerModel(
183
+ attn_num_head_channels,
184
+ in_channels // attn_num_head_channels,
185
+ in_channels=in_channels,
186
+ num_layers=1,
187
+ cross_attention_dim=cross_attention_dim,
188
+ norm_num_groups=resnet_groups,
189
+ use_linear_projection=use_linear_projection,
190
+ upcast_attention=upcast_attention,
191
+ )
192
+ )
193
+ resnets.append(
194
+ ResnetBlockPseudo3D(
195
+ in_channels=in_channels,
196
+ out_channels=in_channels,
197
+ temb_channels=temb_channels,
198
+ eps=resnet_eps,
199
+ groups=resnet_groups,
200
+ dropout=dropout,
201
+ time_embedding_norm=resnet_time_scale_shift,
202
+ non_linearity=resnet_act_fn,
203
+ output_scale_factor=output_scale_factor,
204
+ pre_norm=resnet_pre_norm,
205
+ )
206
+ )
207
+
208
+ self.attentions = nn.ModuleList(attentions)
209
+ self.resnets = nn.ModuleList(resnets)
210
+
211
+ def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None):
212
+ # TODO(Patrick, William) - attention_mask is currently not used. Implement once used
213
+ hidden_states = self.resnets[0](hidden_states, temb)
214
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
215
+ hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
216
+ hidden_states = resnet(hidden_states, temb)
217
+
218
+ return hidden_states
219
+
220
+
221
+ class CrossAttnDownBlockPseudo3D(nn.Module):
222
+ def __init__(
223
+ self,
224
+ in_channels: int,
225
+ out_channels: int,
226
+ temb_channels: int,
227
+ dropout: float = 0.0,
228
+ num_layers: int = 1,
229
+ resnet_eps: float = 1e-6,
230
+ resnet_time_scale_shift: str = "default",
231
+ resnet_act_fn: str = "swish",
232
+ resnet_groups: int = 32,
233
+ resnet_pre_norm: bool = True,
234
+ attn_num_head_channels=1,
235
+ cross_attention_dim=1280,
236
+ output_scale_factor=1.0,
237
+ downsample_padding=1,
238
+ add_downsample=True,
239
+ dual_cross_attention=False,
240
+ use_linear_projection=False,
241
+ only_cross_attention=False,
242
+ upcast_attention=False,
243
+ ):
244
+ super().__init__()
245
+ resnets = []
246
+ attentions = []
247
+
248
+ self.has_cross_attention = True
249
+ self.attn_num_head_channels = attn_num_head_channels
250
+
251
+ for i in range(num_layers):
252
+ in_channels = in_channels if i == 0 else out_channels
253
+ resnets.append(
254
+ ResnetBlockPseudo3D(
255
+ in_channels=in_channels,
256
+ out_channels=out_channels,
257
+ temb_channels=temb_channels,
258
+ eps=resnet_eps,
259
+ groups=resnet_groups,
260
+ dropout=dropout,
261
+ time_embedding_norm=resnet_time_scale_shift,
262
+ non_linearity=resnet_act_fn,
263
+ output_scale_factor=output_scale_factor,
264
+ pre_norm=resnet_pre_norm,
265
+ )
266
+ )
267
+ if dual_cross_attention:
268
+ raise NotImplementedError
269
+ attentions.append(
270
+ SpatioTemporalTransformerModel(
271
+ attn_num_head_channels,
272
+ out_channels // attn_num_head_channels,
273
+ in_channels=out_channels,
274
+ num_layers=1,
275
+ cross_attention_dim=cross_attention_dim,
276
+ norm_num_groups=resnet_groups,
277
+ use_linear_projection=use_linear_projection,
278
+ only_cross_attention=only_cross_attention,
279
+ upcast_attention=upcast_attention,
280
+ )
281
+ )
282
+ self.attentions = nn.ModuleList(attentions)
283
+ self.resnets = nn.ModuleList(resnets)
284
+
285
+ if add_downsample:
286
+ self.downsamplers = nn.ModuleList(
287
+ [
288
+ DownsamplePseudo3D(
289
+ out_channels,
290
+ use_conv=True,
291
+ out_channels=out_channels,
292
+ padding=downsample_padding,
293
+ name="op",
294
+ )
295
+ ]
296
+ )
297
+ else:
298
+ self.downsamplers = None
299
+
300
+ self.gradient_checkpointing = False
301
+
302
+ def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None):
303
+ # TODO(Patrick, William) - attention mask is not used
304
+ output_states = ()
305
+
306
+ for resnet, attn in zip(self.resnets, self.attentions):
307
+ if self.training and self.gradient_checkpointing:
308
+
309
+ def create_custom_forward(module, return_dict=None):
310
+ def custom_forward(*inputs):
311
+ if return_dict is not None:
312
+ return module(*inputs, return_dict=return_dict)
313
+ else:
314
+ return module(*inputs)
315
+
316
+ return custom_forward
317
+
318
+ hidden_states = torch.utils.checkpoint.checkpoint(
319
+ create_custom_forward(resnet), hidden_states, temb
320
+ )
321
+ hidden_states = torch.utils.checkpoint.checkpoint(
322
+ create_custom_forward(attn, return_dict=False),
323
+ hidden_states,
324
+ encoder_hidden_states,
325
+ )[0]
326
+ else:
327
+ hidden_states = resnet(hidden_states, temb)
328
+ hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
329
+
330
+ output_states += (hidden_states,)
331
+
332
+ if self.downsamplers is not None:
333
+ for downsampler in self.downsamplers:
334
+ hidden_states = downsampler(hidden_states)
335
+
336
+ output_states += (hidden_states,)
337
+
338
+ return hidden_states, output_states
339
+
340
+
341
+ class DownBlockPseudo3D(nn.Module):
342
+ def __init__(
343
+ self,
344
+ in_channels: int,
345
+ out_channels: int,
346
+ temb_channels: int,
347
+ dropout: float = 0.0,
348
+ num_layers: int = 1,
349
+ resnet_eps: float = 1e-6,
350
+ resnet_time_scale_shift: str = "default",
351
+ resnet_act_fn: str = "swish",
352
+ resnet_groups: int = 32,
353
+ resnet_pre_norm: bool = True,
354
+ output_scale_factor=1.0,
355
+ add_downsample=True,
356
+ downsample_padding=1,
357
+ ):
358
+ super().__init__()
359
+ resnets = []
360
+
361
+ for i in range(num_layers):
362
+ in_channels = in_channels if i == 0 else out_channels
363
+ resnets.append(
364
+ ResnetBlockPseudo3D(
365
+ in_channels=in_channels,
366
+ out_channels=out_channels,
367
+ temb_channels=temb_channels,
368
+ eps=resnet_eps,
369
+ groups=resnet_groups,
370
+ dropout=dropout,
371
+ time_embedding_norm=resnet_time_scale_shift,
372
+ non_linearity=resnet_act_fn,
373
+ output_scale_factor=output_scale_factor,
374
+ pre_norm=resnet_pre_norm,
375
+ )
376
+ )
377
+
378
+ self.resnets = nn.ModuleList(resnets)
379
+
380
+ if add_downsample:
381
+ self.downsamplers = nn.ModuleList(
382
+ [
383
+ DownsamplePseudo3D(
384
+ out_channels,
385
+ use_conv=True,
386
+ out_channels=out_channels,
387
+ padding=downsample_padding,
388
+ name="op",
389
+ )
390
+ ]
391
+ )
392
+ else:
393
+ self.downsamplers = None
394
+
395
+ self.gradient_checkpointing = False
396
+
397
+ def forward(self, hidden_states, temb=None):
398
+ output_states = ()
399
+
400
+ for resnet in self.resnets:
401
+ if self.training and self.gradient_checkpointing:
402
+
403
+ def create_custom_forward(module):
404
+ def custom_forward(*inputs):
405
+ return module(*inputs)
406
+
407
+ return custom_forward
408
+
409
+ hidden_states = torch.utils.checkpoint.checkpoint(
410
+ create_custom_forward(resnet), hidden_states, temb
411
+ )
412
+ else:
413
+ hidden_states = resnet(hidden_states, temb)
414
+
415
+ output_states += (hidden_states,)
416
+
417
+ if self.downsamplers is not None:
418
+ for downsampler in self.downsamplers:
419
+ hidden_states = downsampler(hidden_states)
420
+
421
+ output_states += (hidden_states,)
422
+
423
+ return hidden_states, output_states
424
+
425
+
426
+ class CrossAttnUpBlockPseudo3D(nn.Module):
427
+ def __init__(
428
+ self,
429
+ in_channels: int,
430
+ out_channels: int,
431
+ prev_output_channel: int,
432
+ temb_channels: int,
433
+ dropout: float = 0.0,
434
+ num_layers: int = 1,
435
+ resnet_eps: float = 1e-6,
436
+ resnet_time_scale_shift: str = "default",
437
+ resnet_act_fn: str = "swish",
438
+ resnet_groups: int = 32,
439
+ resnet_pre_norm: bool = True,
440
+ attn_num_head_channels=1,
441
+ cross_attention_dim=1280,
442
+ output_scale_factor=1.0,
443
+ add_upsample=True,
444
+ dual_cross_attention=False,
445
+ use_linear_projection=False,
446
+ only_cross_attention=False,
447
+ upcast_attention=False,
448
+ ):
449
+ super().__init__()
450
+ resnets = []
451
+ attentions = []
452
+
453
+ self.has_cross_attention = True
454
+ self.attn_num_head_channels = attn_num_head_channels
455
+
456
+ for i in range(num_layers):
457
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
458
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
459
+
460
+ resnets.append(
461
+ ResnetBlockPseudo3D(
462
+ in_channels=resnet_in_channels + res_skip_channels,
463
+ out_channels=out_channels,
464
+ temb_channels=temb_channels,
465
+ eps=resnet_eps,
466
+ groups=resnet_groups,
467
+ dropout=dropout,
468
+ time_embedding_norm=resnet_time_scale_shift,
469
+ non_linearity=resnet_act_fn,
470
+ output_scale_factor=output_scale_factor,
471
+ pre_norm=resnet_pre_norm,
472
+ )
473
+ )
474
+ if dual_cross_attention:
475
+ raise NotImplementedError
476
+ attentions.append(
477
+ SpatioTemporalTransformerModel(
478
+ attn_num_head_channels,
479
+ out_channels // attn_num_head_channels,
480
+ in_channels=out_channels,
481
+ num_layers=1,
482
+ cross_attention_dim=cross_attention_dim,
483
+ norm_num_groups=resnet_groups,
484
+ use_linear_projection=use_linear_projection,
485
+ only_cross_attention=only_cross_attention,
486
+ upcast_attention=upcast_attention,
487
+ )
488
+ )
489
+ self.attentions = nn.ModuleList(attentions)
490
+ self.resnets = nn.ModuleList(resnets)
491
+
492
+ if add_upsample:
493
+ self.upsamplers = nn.ModuleList(
494
+ [UpsamplePseudo3D(out_channels, use_conv=True, out_channels=out_channels)]
495
+ )
496
+ else:
497
+ self.upsamplers = None
498
+
499
+ self.gradient_checkpointing = False
500
+
501
+ def forward(
502
+ self,
503
+ hidden_states,
504
+ res_hidden_states_tuple,
505
+ temb=None,
506
+ encoder_hidden_states=None,
507
+ upsample_size=None,
508
+ attention_mask=None,
509
+ ):
510
+ # TODO(Patrick, William) - attention mask is not used
511
+ for resnet, attn in zip(self.resnets, self.attentions):
512
+ # pop res hidden states
513
+ res_hidden_states = res_hidden_states_tuple[-1]
514
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
515
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
516
+
517
+ if self.training and self.gradient_checkpointing:
518
+
519
+ def create_custom_forward(module, return_dict=None):
520
+ def custom_forward(*inputs):
521
+ if return_dict is not None:
522
+ return module(*inputs, return_dict=return_dict)
523
+ else:
524
+ return module(*inputs)
525
+
526
+ return custom_forward
527
+
528
+ hidden_states = torch.utils.checkpoint.checkpoint(
529
+ create_custom_forward(resnet), hidden_states, temb
530
+ )
531
+ hidden_states = torch.utils.checkpoint.checkpoint(
532
+ create_custom_forward(attn, return_dict=False),
533
+ hidden_states,
534
+ encoder_hidden_states,
535
+ )[0]
536
+ else:
537
+ hidden_states = resnet(hidden_states, temb)
538
+ hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
539
+
540
+ if self.upsamplers is not None:
541
+ for upsampler in self.upsamplers:
542
+ hidden_states = upsampler(hidden_states, upsample_size)
543
+
544
+ return hidden_states
545
+
546
+
547
+ class UpBlockPseudo3D(nn.Module):
548
+ def __init__(
549
+ self,
550
+ in_channels: int,
551
+ prev_output_channel: int,
552
+ out_channels: int,
553
+ temb_channels: int,
554
+ dropout: float = 0.0,
555
+ num_layers: int = 1,
556
+ resnet_eps: float = 1e-6,
557
+ resnet_time_scale_shift: str = "default",
558
+ resnet_act_fn: str = "swish",
559
+ resnet_groups: int = 32,
560
+ resnet_pre_norm: bool = True,
561
+ output_scale_factor=1.0,
562
+ add_upsample=True,
563
+ ):
564
+ super().__init__()
565
+ resnets = []
566
+
567
+ for i in range(num_layers):
568
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
569
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
570
+
571
+ resnets.append(
572
+ ResnetBlockPseudo3D(
573
+ in_channels=resnet_in_channels + res_skip_channels,
574
+ out_channels=out_channels,
575
+ temb_channels=temb_channels,
576
+ eps=resnet_eps,
577
+ groups=resnet_groups,
578
+ dropout=dropout,
579
+ time_embedding_norm=resnet_time_scale_shift,
580
+ non_linearity=resnet_act_fn,
581
+ output_scale_factor=output_scale_factor,
582
+ pre_norm=resnet_pre_norm,
583
+ )
584
+ )
585
+
586
+ self.resnets = nn.ModuleList(resnets)
587
+
588
+ if add_upsample:
589
+ self.upsamplers = nn.ModuleList(
590
+ [UpsamplePseudo3D(out_channels, use_conv=True, out_channels=out_channels)]
591
+ )
592
+ else:
593
+ self.upsamplers = None
594
+
595
+ self.gradient_checkpointing = False
596
+
597
+ def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
598
+ for resnet in self.resnets:
599
+ # pop res hidden states
600
+ res_hidden_states = res_hidden_states_tuple[-1]
601
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
602
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
603
+
604
+ if self.training and self.gradient_checkpointing:
605
+
606
+ def create_custom_forward(module):
607
+ def custom_forward(*inputs):
608
+ return module(*inputs)
609
+
610
+ return custom_forward
611
+
612
+ hidden_states = torch.utils.checkpoint.checkpoint(
613
+ create_custom_forward(resnet), hidden_states, temb
614
+ )
615
+ else:
616
+ hidden_states = resnet(hidden_states, temb)
617
+
618
+ if self.upsamplers is not None:
619
+ for upsampler in self.upsamplers:
620
+ hidden_states = upsampler(hidden_states, upsample_size)
621
+
622
+ return hidden_states
model/video_diffusion/models/unet_3d_blocks_control.py ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import torch
16
+ from torch import nn
17
+ from .attention import SpatioTemporalTransformerModel
18
+ from .resnet import DownsamplePseudo3D, ResnetBlockPseudo3D, UpsamplePseudo3D
19
+ import glob
20
+ import json
21
+ from dataclasses import dataclass
22
+ from typing import List, Optional, Tuple, Union
23
+ import torch
24
+ import torch.nn as nn
25
+ import torch.utils.checkpoint
26
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
27
+ from diffusers.models.modeling_utils import ModelMixin
28
+ from diffusers.utils import BaseOutput, logging
29
+ from diffusers.models.embeddings import TimestepEmbedding, Timesteps
30
+ from .unet_3d_blocks import (
31
+ CrossAttnDownBlockPseudo3D,
32
+ CrossAttnUpBlockPseudo3D,
33
+ DownBlockPseudo3D,
34
+ UNetMidBlockPseudo3DCrossAttn,
35
+ UpBlockPseudo3D,
36
+ get_down_block,
37
+ get_up_block,
38
+ )
39
+ from .resnet import PseudoConv3d
40
+ from diffusers.models.cross_attention import AttnProcessor
41
+ from typing import Dict
42
+
43
+
44
+
45
+ def set_zero_parameters(module):
46
+ for p in module.parameters():
47
+ p.detach().zero_()
48
+ return module
49
+
50
+ # ControlNet: Zero Convolution
51
+ def zero_conv(channels):
52
+ return set_zero_parameters(PseudoConv3d(channels, channels, 1, padding=0))
53
+
54
+ class ControlNetInputHintBlock(nn.Module):
55
+ def __init__(self, hint_channels: int = 3, channels: int = 320):
56
+ super().__init__()
57
+ # Layer configurations are from reference implementation.
58
+ self.input_hint_block = nn.Sequential(
59
+ PseudoConv3d(hint_channels, 16, 3, padding=1),
60
+ nn.SiLU(),
61
+ PseudoConv3d(16, 16, 3, padding=1),
62
+ nn.SiLU(),
63
+ PseudoConv3d(16, 32, 3, padding=1, stride=2),
64
+ nn.SiLU(),
65
+ PseudoConv3d(32, 32, 3, padding=1),
66
+ nn.SiLU(),
67
+ PseudoConv3d(32, 96, 3, padding=1, stride=2),
68
+ nn.SiLU(),
69
+ PseudoConv3d(96, 96, 3, padding=1),
70
+ nn.SiLU(),
71
+ PseudoConv3d(96, 256, 3, padding=1, stride=2),
72
+ nn.SiLU(),
73
+ set_zero_parameters(PseudoConv3d(256, channels, 3, padding=1)),
74
+ )
75
+ def forward(self, hint: torch.Tensor):
76
+ return self.input_hint_block(hint)
77
+
78
+
79
+ class ControlNetPseudoZeroConv3dBlock(nn.Module):
80
+ def __init__(
81
+ self,
82
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
83
+ down_block_types: Tuple[str] = (
84
+ "CrossAttnDownBlockPseudo3D",
85
+ "CrossAttnDownBlockPseudo3D",
86
+ "CrossAttnDownBlockPseudo3D",
87
+ "DownBlockPseudo3D",
88
+ ),
89
+ layers_per_block: int = 2,
90
+ ):
91
+ super().__init__()
92
+ self.input_zero_conv = zero_conv(block_out_channels[0])
93
+ zero_convs = []
94
+ for i, down_block_type in enumerate(down_block_types):
95
+ output_channel = block_out_channels[i]
96
+ is_final_block = i == len(block_out_channels) - 1
97
+ for _ in range(layers_per_block):
98
+ zero_convs.append(zero_conv(output_channel))
99
+ if not is_final_block:
100
+ zero_convs.append(zero_conv(output_channel))
101
+ self.zero_convs = nn.ModuleList(zero_convs)
102
+ self.mid_zero_conv = zero_conv(block_out_channels[-1])
103
+
104
+ def forward(
105
+ self,
106
+ down_block_res_samples: List[torch.Tensor],
107
+ mid_block_sample: torch.Tensor,
108
+ ) -> List[torch.Tensor]:
109
+ outputs = []
110
+ outputs.append(self.input_zero_conv(down_block_res_samples[0]))
111
+ for res_sample, zero_conv in zip(down_block_res_samples[1:], self.zero_convs):
112
+ outputs.append(zero_conv(res_sample))
113
+ outputs.append(self.mid_zero_conv(mid_block_sample))
114
+ return outputs
115
+
116
+
model/video_diffusion/models/unet_3d_condition.py ADDED
@@ -0,0 +1,571 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import os
16
+ import glob
17
+ import json
18
+ from dataclasses import dataclass
19
+ from typing import List, Optional, Tuple, Union
20
+
21
+ import torch
22
+ import torch.nn as nn
23
+ import torch.utils.checkpoint
24
+
25
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
26
+ from diffusers.models.modeling_utils import ModelMixin
27
+ from diffusers.utils import BaseOutput, logging
28
+ from diffusers.models.embeddings import TimestepEmbedding, Timesteps
29
+ from .unet_3d_blocks import (
30
+ CrossAttnDownBlockPseudo3D,
31
+ CrossAttnUpBlockPseudo3D,
32
+ DownBlockPseudo3D,
33
+ UNetMidBlockPseudo3DCrossAttn,
34
+ UpBlockPseudo3D,
35
+ get_down_block,
36
+ get_up_block,
37
+ )
38
+ from .resnet import PseudoConv3d
39
+ from diffusers.models.cross_attention import AttnProcessor
40
+ from typing import Dict
41
+
42
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
43
+
44
+
45
+ @dataclass
46
+ class UNetPseudo3DConditionOutput(BaseOutput):
47
+ sample: torch.FloatTensor
48
+
49
+
50
+ class UNetPseudo3DConditionModel(ModelMixin, ConfigMixin):
51
+ """
52
+ 这里把原来2D Unet的 2D卷积全换成新定义的PseudoConv3d。并且定义了从2D卷积继承的模型参数。
53
+ """
54
+ _supports_gradient_checkpointing = True
55
+
56
+ @register_to_config
57
+ def __init__(
58
+ self,
59
+ sample_size: Optional[int] = None,
60
+ in_channels: int = 4,
61
+ out_channels: int = 4,
62
+ center_input_sample: bool = False,
63
+ flip_sin_to_cos: bool = True,
64
+ freq_shift: int = 0,
65
+ down_block_types: Tuple[str] = (
66
+ "CrossAttnDownBlockPseudo3D",
67
+ "CrossAttnDownBlockPseudo3D",
68
+ "CrossAttnDownBlockPseudo3D",
69
+ "DownBlockPseudo3D",
70
+ ),
71
+ mid_block_type: str = "UNetMidBlockPseudo3DCrossAttn",
72
+ up_block_types: Tuple[str] = (
73
+ "UpBlockPseudo3D",
74
+ "CrossAttnUpBlockPseudo3D",
75
+ "CrossAttnUpBlockPseudo3D",
76
+ "CrossAttnUpBlockPseudo3D",
77
+ ),
78
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
79
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
80
+ layers_per_block: int = 2,
81
+ downsample_padding: int = 1,
82
+ mid_block_scale_factor: float = 1,
83
+ act_fn: str = "silu",
84
+ norm_num_groups: int = 32,
85
+ norm_eps: float = 1e-5,
86
+ cross_attention_dim: int = 1280,
87
+ attention_head_dim: Union[int, Tuple[int]] = 8,
88
+ dual_cross_attention: bool = False,
89
+ use_linear_projection: bool = False,
90
+ fps_embed_type: Optional[str] = None,
91
+ num_fps_embeds: Optional[int] = None,
92
+ upcast_attention: bool = False,
93
+ resnet_time_scale_shift: str = "default",
94
+ num_class_embeds=None,
95
+
96
+ ):
97
+ super().__init__()
98
+
99
+
100
+ self.sample_size = sample_size
101
+ time_embed_dim = block_out_channels[0] * 4
102
+
103
+ # input
104
+ self.conv_in = PseudoConv3d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
105
+
106
+ # time
107
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
108
+ timestep_input_dim = block_out_channels[0]
109
+
110
+ self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
111
+
112
+ # class embedding
113
+ if fps_embed_type is None and num_fps_embeds is not None:
114
+ self.fps_embedding = nn.Embedding(num_fps_embeds, time_embed_dim)
115
+ elif fps_embed_type == "timestep":
116
+ self.fps_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
117
+ elif fps_embed_type == "identity":
118
+ self.fps_embedding = nn.Identity(time_embed_dim, time_embed_dim)
119
+ else:
120
+ self.fps_embedding = None
121
+
122
+ self.down_blocks = nn.ModuleList([])
123
+ self.mid_block = None
124
+ self.up_blocks = nn.ModuleList([])
125
+
126
+ if isinstance(only_cross_attention, bool):
127
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
128
+
129
+ if isinstance(attention_head_dim, int):
130
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
131
+
132
+ # down
133
+ output_channel = block_out_channels[0]
134
+ for i, down_block_type in enumerate(down_block_types):
135
+ input_channel = output_channel
136
+ output_channel = block_out_channels[i]
137
+ is_final_block = i == len(block_out_channels) - 1
138
+
139
+ down_block = get_down_block(
140
+ down_block_type,
141
+ num_layers=layers_per_block,
142
+ in_channels=input_channel,
143
+ out_channels=output_channel,
144
+ temb_channels=time_embed_dim,
145
+ add_downsample=not is_final_block,
146
+ resnet_eps=norm_eps,
147
+ resnet_act_fn=act_fn,
148
+ resnet_groups=norm_num_groups,
149
+ cross_attention_dim=cross_attention_dim,
150
+ attn_num_head_channels=attention_head_dim[i],
151
+ downsample_padding=downsample_padding,
152
+ dual_cross_attention=dual_cross_attention,
153
+ use_linear_projection=use_linear_projection,
154
+ only_cross_attention=only_cross_attention[i],
155
+ upcast_attention=upcast_attention,
156
+ resnet_time_scale_shift=resnet_time_scale_shift,
157
+ )
158
+ self.down_blocks.append(down_block)
159
+
160
+ # mid
161
+ if mid_block_type == "UNetMidBlockPseudo3DCrossAttn":
162
+ self.mid_block = UNetMidBlockPseudo3DCrossAttn(
163
+ in_channels=block_out_channels[-1],
164
+ temb_channels=time_embed_dim,
165
+ resnet_eps=norm_eps,
166
+ resnet_act_fn=act_fn,
167
+ output_scale_factor=mid_block_scale_factor,
168
+ resnet_time_scale_shift=resnet_time_scale_shift,
169
+ cross_attention_dim=cross_attention_dim,
170
+ attn_num_head_channels=attention_head_dim[-1],
171
+ resnet_groups=norm_num_groups,
172
+ dual_cross_attention=dual_cross_attention,
173
+ use_linear_projection=use_linear_projection,
174
+ upcast_attention=upcast_attention,
175
+ )
176
+ else:
177
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
178
+
179
+ # count how many layers upsample the images
180
+ self.num_upsamplers = 0
181
+
182
+ # up
183
+ reversed_block_out_channels = list(reversed(block_out_channels))
184
+ reversed_attention_head_dim = list(reversed(attention_head_dim))
185
+ only_cross_attention = list(reversed(only_cross_attention))
186
+ output_channel = reversed_block_out_channels[0]
187
+ for i, up_block_type in enumerate(up_block_types):
188
+ is_final_block = i == len(block_out_channels) - 1
189
+
190
+ prev_output_channel = output_channel
191
+ output_channel = reversed_block_out_channels[i]
192
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
193
+
194
+ # add upsample block for all BUT final layer
195
+ if not is_final_block:
196
+ add_upsample = True
197
+ self.num_upsamplers += 1
198
+ else:
199
+ add_upsample = False
200
+
201
+ up_block = get_up_block(
202
+ up_block_type,
203
+ num_layers=layers_per_block + 1,
204
+ in_channels=input_channel,
205
+ out_channels=output_channel,
206
+ prev_output_channel=prev_output_channel,
207
+ temb_channels=time_embed_dim,
208
+ add_upsample=add_upsample,
209
+ resnet_eps=norm_eps,
210
+ resnet_act_fn=act_fn,
211
+ resnet_groups=norm_num_groups,
212
+ cross_attention_dim=cross_attention_dim,
213
+ attn_num_head_channels=reversed_attention_head_dim[i],
214
+ dual_cross_attention=dual_cross_attention,
215
+ use_linear_projection=use_linear_projection,
216
+ only_cross_attention=only_cross_attention[i],
217
+ upcast_attention=upcast_attention,
218
+ resnet_time_scale_shift=resnet_time_scale_shift,
219
+ )
220
+ self.up_blocks.append(up_block)
221
+ prev_output_channel = output_channel
222
+
223
+ # out
224
+ self.conv_norm_out = nn.GroupNorm(
225
+ num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
226
+ )
227
+ self.conv_act = nn.SiLU()
228
+ self.conv_out = PseudoConv3d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
229
+ @property
230
+ def attn_processors(self) -> Dict[str, AttnProcessor]:
231
+ r"""
232
+ Returns:
233
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
234
+ indexed by its weight name.
235
+ """
236
+ # set recursively
237
+ processors = {}
238
+
239
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttnProcessor]):
240
+ if hasattr(module, "set_processor"):
241
+ processors[f"{name}.processor"] = module.processor
242
+
243
+ for sub_name, child in module.named_children():
244
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
245
+
246
+ return processors
247
+
248
+ for name, module in self.named_children():
249
+ fn_recursive_add_processors(name, module, processors)
250
+
251
+ return processors
252
+
253
+ def set_attn_processor(self, processor: Union[AttnProcessor, Dict[str, AttnProcessor]]):
254
+ r"""
255
+ Parameters:
256
+ `processor (`dict` of `AttnProcessor` or `AttnProcessor`):
257
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
258
+ of **all** `CrossAttention` layers.
259
+ In case `processor` is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainablae attention processors.:
260
+
261
+ """
262
+ count = len(self.attn_processors.keys())
263
+
264
+ if isinstance(processor, dict) and len(processor) != count:
265
+ raise ValueError(
266
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
267
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
268
+ )
269
+
270
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
271
+ if hasattr(module, "set_processor"):
272
+ if not isinstance(processor, dict):
273
+ module.set_processor(processor)
274
+ else:
275
+ module.set_processor(processor.pop(f"{name}.processor"))
276
+
277
+ for sub_name, child in module.named_children():
278
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
279
+
280
+ for name, module in self.named_children():
281
+ fn_recursive_attn_processor(name, module, processor)
282
+
283
+
284
+ def set_attention_slice(self, slice_size):
285
+ r"""
286
+ Enable sliced attention computation.
287
+
288
+ When this option is enabled, the attention module will split the input tensor in slices, to compute attention
289
+ in several steps. This is useful to save some memory in exchange for a small speed decrease.
290
+
291
+ Args:
292
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
293
+ When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
294
+ `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
295
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
296
+ must be a multiple of `slice_size`.
297
+ """
298
+ sliceable_head_dims = []
299
+
300
+ def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
301
+ if hasattr(module, "set_attention_slice"):
302
+ sliceable_head_dims.append(module.sliceable_head_dim)
303
+
304
+ for child in module.children():
305
+ fn_recursive_retrieve_slicable_dims(child)
306
+
307
+ # retrieve number of attention layers
308
+ for module in self.children():
309
+ fn_recursive_retrieve_slicable_dims(module)
310
+
311
+ num_slicable_layers = len(sliceable_head_dims)
312
+
313
+ if slice_size == "auto":
314
+ # half the attention head size is usually a good trade-off between
315
+ # speed and memory
316
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
317
+ elif slice_size == "max":
318
+ # make smallest slice possible
319
+ slice_size = num_slicable_layers * [1]
320
+
321
+ slice_size = (
322
+ num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
323
+ )
324
+
325
+ if len(slice_size) != len(sliceable_head_dims):
326
+ raise ValueError(
327
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
328
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
329
+ )
330
+
331
+ for i in range(len(slice_size)):
332
+ size = slice_size[i]
333
+ dim = sliceable_head_dims[i]
334
+ if size is not None and size > dim:
335
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
336
+
337
+ # Recursively walk through all the children.
338
+ # Any children which exposes the set_attention_slice method
339
+ # gets the message
340
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
341
+ if hasattr(module, "set_attention_slice"):
342
+ module.set_attention_slice(slice_size.pop())
343
+
344
+ for child in module.children():
345
+ fn_recursive_set_attention_slice(child, slice_size)
346
+
347
+ reversed_slice_size = list(reversed(slice_size))
348
+ for module in self.children():
349
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
350
+
351
+ def _set_gradient_checkpointing(self, module, value=False):
352
+ if isinstance(
353
+ module,
354
+ (CrossAttnDownBlockPseudo3D, DownBlockPseudo3D, CrossAttnUpBlockPseudo3D, UpBlockPseudo3D),
355
+ ):
356
+ module.gradient_checkpointing = value
357
+
358
+ def forward(
359
+ self,
360
+ sample: torch.FloatTensor,
361
+ timestep: Union[torch.Tensor, float, int],
362
+ encoder_hidden_states: torch.Tensor,
363
+ fps_labels: Optional[torch.Tensor] = None,
364
+ attention_mask: Optional[torch.Tensor] = None,
365
+ cross_attention_kwargs=None,
366
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
367
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
368
+ return_dict: bool = True,
369
+ ) -> Union[UNetPseudo3DConditionOutput, Tuple]:
370
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
371
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
372
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
373
+ # on the fly if necessary.
374
+ default_overall_up_factor = 2**self.num_upsamplers
375
+
376
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
377
+ forward_upsample_size = False
378
+ upsample_size = None
379
+
380
+ if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
381
+ logger.info("Forward upsample size to force interpolation output size.")
382
+ forward_upsample_size = True
383
+
384
+ # prepare attention_mask
385
+ if attention_mask is not None:
386
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
387
+ attention_mask = attention_mask.unsqueeze(1)
388
+
389
+ # 0. center input if necessary
390
+ if self.config.center_input_sample:
391
+ sample = 2 * sample - 1.0
392
+
393
+ # 1. time
394
+ timesteps = timestep
395
+ if not torch.is_tensor(timesteps):
396
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
397
+ # This would be a good case for the `match` statement (Python 3.10+)
398
+ is_mps = sample.device.type == "mps"
399
+ if isinstance(timestep, float):
400
+ dtype = torch.float32 if is_mps else torch.float64
401
+ else:
402
+ dtype = torch.int32 if is_mps else torch.int64
403
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
404
+ elif len(timesteps.shape) == 0:
405
+ timesteps = timesteps[None].to(sample.device)
406
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
407
+ timesteps = timesteps.expand(sample.shape[0])
408
+
409
+ t_emb = self.time_proj(timesteps)
410
+ # timesteps does not contain any weights and will always return f32 tensors
411
+ # but time_embedding might actually be running in fp16. so we need to cast here.
412
+ # there might be better ways to encapsulate this.
413
+ t_emb = t_emb.to(dtype=self.dtype)
414
+ emb = self.time_embedding(t_emb)
415
+
416
+ if self.fps_embedding is not None:
417
+ if fps_labels is None:
418
+ raise ValueError("fps_labels should be provided when num_fps_embeds > 0")
419
+
420
+ if self.config.fps_embed_type == "timestep":
421
+ fps_labels = self.time_proj(fps_labels) # 和timesteps共用,都是sin embedding?这里的weight不更新的。
422
+
423
+ # 这里和上面timesteps does not contain any weights and will always return f32 tensors的bug一样。需要先cast过去,不然多机多卡就有问题了。
424
+ fps_labels = fps_labels.to(dtype=self.dtype)
425
+ class_emb = self.fps_embedding(fps_labels)
426
+
427
+ emb = emb + class_emb
428
+
429
+ # 2. pre-process
430
+ sample = self.conv_in(sample)
431
+
432
+ # 3. down
433
+ down_block_res_samples = (sample,)
434
+ for downsample_block in self.down_blocks:
435
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
436
+ sample, res_samples = downsample_block(
437
+ hidden_states=sample,
438
+ temb=emb,
439
+ encoder_hidden_states=encoder_hidden_states,
440
+ attention_mask=attention_mask,
441
+ )
442
+ else:
443
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
444
+
445
+ down_block_res_samples += res_samples
446
+
447
+ if down_block_additional_residuals is not None:
448
+ new_down_block_res_samples = ()
449
+
450
+ for down_block_res_sample, down_block_additional_residual in zip(
451
+ down_block_res_samples, down_block_additional_residuals
452
+ ):
453
+ down_block_res_sample = down_block_res_sample + down_block_additional_residual
454
+ new_down_block_res_samples += (down_block_res_sample,)
455
+
456
+ down_block_res_samples = new_down_block_res_samples
457
+
458
+ # 4. mid
459
+ sample = self.mid_block(
460
+ sample, emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
461
+ )
462
+ if mid_block_additional_residual is not None:
463
+ sample = sample + mid_block_additional_residual
464
+
465
+ # 5. up
466
+ for i, upsample_block in enumerate(self.up_blocks):
467
+ is_final_block = i == len(self.up_blocks) - 1
468
+
469
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
470
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
471
+
472
+ # if we have not reached the final block and need to forward the
473
+ # upsample size, we do it here
474
+ if not is_final_block and forward_upsample_size:
475
+ upsample_size = down_block_res_samples[-1].shape[2:]
476
+
477
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
478
+ sample = upsample_block(
479
+ hidden_states=sample,
480
+ temb=emb,
481
+ res_hidden_states_tuple=res_samples,
482
+ encoder_hidden_states=encoder_hidden_states,
483
+ upsample_size=upsample_size,
484
+ attention_mask=attention_mask,
485
+ )
486
+ else:
487
+ sample = upsample_block(
488
+ hidden_states=sample,
489
+ temb=emb,
490
+ res_hidden_states_tuple=res_samples,
491
+ upsample_size=upsample_size,
492
+ )
493
+ # 6. post-process
494
+ sample = self.conv_norm_out(sample)
495
+ sample = self.conv_act(sample)
496
+ sample = self.conv_out(sample)
497
+
498
+ if not return_dict:
499
+ return (sample,)
500
+
501
+ return UNetPseudo3DConditionOutput(sample=sample)
502
+
503
+ @classmethod
504
+ def from_2d_model(cls, model_path, condition_on_fps=False):
505
+ '''
506
+ load a 2d model and convert it to a pseudo 3d model
507
+ '''
508
+ config_path = os.path.join(model_path, "config.json")
509
+ if not os.path.isfile(config_path):
510
+ raise RuntimeError(f"{config_path} does not exist")
511
+ with open(config_path, "r") as f:
512
+ config = json.load(f)
513
+
514
+ config.pop("_class_name")
515
+ config.pop("_diffusers_version")
516
+
517
+ block_replacer = {
518
+ "CrossAttnDownBlock2D": "CrossAttnDownBlockPseudo3D",
519
+ "DownBlock2D": "DownBlockPseudo3D",
520
+ "UpBlock2D": "UpBlockPseudo3D",
521
+ "CrossAttnUpBlock2D": "CrossAttnUpBlockPseudo3D",
522
+ }
523
+
524
+ def convert_2d_to_3d_block(block):
525
+ return block_replacer[block] if block in block_replacer else block
526
+
527
+ config["down_block_types"] = [
528
+ convert_2d_to_3d_block(block) for block in config["down_block_types"]
529
+ ]
530
+ config["up_block_types"] = [convert_2d_to_3d_block(block) for block in config["up_block_types"]]
531
+
532
+ if condition_on_fps:
533
+ # config["num_fps_embeds"] = 60 # 这个在 trainable embeding时候才需要~
534
+ config["fps_embed_type"] = "timestep" # 和timestep保持一致的type。
535
+
536
+
537
+ model = cls(**config) # 调用自身(init), 传入config参数全换成3d的setting
538
+
539
+ state_dict_path_condidates = glob.glob(os.path.join(model_path, "*.bin"))
540
+ if state_dict_path_condidates:
541
+ state_dict = torch.load(state_dict_path_condidates[0], map_location="cpu")
542
+ model.load_2d_state_dict(state_dict=state_dict)
543
+
544
+ return model
545
+
546
+ def load_2d_state_dict(self, state_dict, **kwargs):
547
+ '''
548
+ 2D 部分的参数名完全不变。
549
+ '''
550
+ state_dict_3d = self.state_dict()
551
+
552
+ for k, v in state_dict.items():
553
+ if k not in state_dict_3d:
554
+ raise KeyError(f"2d state_dict key {k} does not exist in 3d model")
555
+ elif v.shape != state_dict_3d[k].shape:
556
+ raise ValueError(f"state_dict shape mismatch, 2d {v.shape}, 3d {state_dict_3d[k].shape}")
557
+
558
+ for k, v in state_dict_3d.items():
559
+ if "_temporal" in k:
560
+ continue
561
+ if "gamma" in k:
562
+ continue
563
+
564
+ if k not in state_dict:
565
+ if "fps_embedding" in k:
566
+ # 忽略检查fps_embedding
567
+ continue
568
+ raise KeyError(f"3d state_dict key {k} does not exist in 2d model")
569
+
570
+ state_dict_3d.update(state_dict)
571
+ self.load_state_dict(state_dict_3d, **kwargs)
model/video_diffusion/pipelines/__init__.py ADDED
File without changes
model/video_diffusion/pipelines/__pycache__/__init__.cpython-39.pyc ADDED
Binary file (178 Bytes). View file
 
model/video_diffusion/pipelines/__pycache__/pipeline_st_stable_diffusion.cpython-39.pyc ADDED
Binary file (19.3 kB). View file
 
model/video_diffusion/pipelines/__pycache__/pipeline_stable_diffusion_controlnet3d.cpython-39.pyc ADDED
Binary file (12.1 kB). View file
 
model/video_diffusion/pipelines/pipeline_st_stable_diffusion.py ADDED
@@ -0,0 +1,618 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import inspect
16
+ from typing import Callable, List, Optional, Union
17
+
18
+ import torch
19
+ from einops import rearrange
20
+
21
+ from diffusers.utils import is_accelerate_available
22
+ from packaging import version
23
+ from transformers import CLIPTextModel, CLIPTokenizer
24
+
25
+ from diffusers.configuration_utils import FrozenDict
26
+ from diffusers.models import AutoencoderKL
27
+ from diffusers.pipeline_utils import DiffusionPipeline
28
+ from diffusers.schedulers import (
29
+ DDIMScheduler,
30
+ DPMSolverMultistepScheduler,
31
+ EulerAncestralDiscreteScheduler,
32
+ EulerDiscreteScheduler,
33
+ LMSDiscreteScheduler,
34
+ PNDMScheduler,
35
+ )
36
+ from diffusers.utils import deprecate, logging
37
+ from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
38
+
39
+ from ..models.unet_3d_condition import UNetPseudo3DConditionModel
40
+ import os, importlib
41
+
42
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
43
+
44
+
45
+ class SpatioTemporalStableDiffusionPipeline(DiffusionPipeline):
46
+ r"""
47
+ Pipeline for text-to-video generation using Spatio-Temporal Stable Diffusion.
48
+ 改变了unet的输入, unet换成3d unet, 其他部分完全和原来2D的一致。
49
+ latents的变为 b,c,f,h,w 原来是 b,c,h,w。
50
+ 要用VAE的decoder的时候, 把输入reshape 成 (b f) c h w
51
+ """
52
+ _optional_components = []
53
+
54
+ def __init__(
55
+ self,
56
+ vae: AutoencoderKL,
57
+ text_encoder: CLIPTextModel,
58
+ tokenizer: CLIPTokenizer,
59
+ unet: UNetPseudo3DConditionModel,
60
+ scheduler: Union[
61
+ DDIMScheduler,
62
+ PNDMScheduler,
63
+ LMSDiscreteScheduler,
64
+ EulerDiscreteScheduler,
65
+ EulerAncestralDiscreteScheduler,
66
+ DPMSolverMultistepScheduler,
67
+ ],
68
+ ):
69
+ super().__init__()
70
+
71
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
72
+ deprecation_message = (
73
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
74
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
75
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
76
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
77
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
78
+ " file"
79
+ )
80
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
81
+ new_config = dict(scheduler.config)
82
+ new_config["steps_offset"] = 1
83
+ scheduler._internal_dict = FrozenDict(new_config)
84
+
85
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
86
+ deprecation_message = (
87
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
88
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
89
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
90
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
91
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
92
+ )
93
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
94
+ new_config = dict(scheduler.config)
95
+ new_config["clip_sample"] = False
96
+ scheduler._internal_dict = FrozenDict(new_config)
97
+
98
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
99
+ version.parse(unet.config._diffusers_version).base_version
100
+ ) < version.parse("0.9.0.dev0")
101
+ is_unet_sample_size_less_64 = (
102
+ hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
103
+ )
104
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
105
+ deprecation_message = (
106
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
107
+ " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
108
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
109
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
110
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
111
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
112
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
113
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
114
+ " the `unet/config.json` file"
115
+ )
116
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
117
+ new_config = dict(unet.config)
118
+ new_config["sample_size"] = 64
119
+ unet._internal_dict = FrozenDict(new_config)
120
+
121
+ self.register_modules(
122
+ vae=vae,
123
+ text_encoder=text_encoder,
124
+ tokenizer=tokenizer,
125
+ unet=unet,
126
+ scheduler=scheduler,
127
+ )
128
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
129
+
130
+ def enable_vae_slicing(self):
131
+ r"""
132
+ Enable sliced VAE decoding.
133
+
134
+ When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
135
+ steps. This is useful to save some memory and allow larger batch sizes.
136
+ """
137
+ self.vae.enable_slicing()
138
+
139
+ def disable_vae_slicing(self):
140
+ r"""
141
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
142
+ computing decoding in one step.
143
+ """
144
+ self.vae.disable_slicing()
145
+
146
+ def enable_sequential_cpu_offload(self, gpu_id=0):
147
+ r"""
148
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
149
+ text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
150
+ `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
151
+ """
152
+ if is_accelerate_available():
153
+ from accelerate import cpu_offload
154
+ else:
155
+ raise ImportError("Please install accelerate via `pip install accelerate`")
156
+
157
+ device = torch.device(f"cuda:{gpu_id}")
158
+
159
+ for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
160
+ if cpu_offloaded_model is not None:
161
+ cpu_offload(cpu_offloaded_model, device)
162
+
163
+ @property
164
+ def _execution_device(self):
165
+ r"""
166
+ Returns the device on which the pipeline's models will be executed. After calling
167
+ `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
168
+ hooks.
169
+ """
170
+ if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
171
+ return self.device
172
+ for module in self.unet.modules():
173
+ if (
174
+ hasattr(module, "_hf_hook")
175
+ and hasattr(module._hf_hook, "execution_device")
176
+ and module._hf_hook.execution_device is not None
177
+ ):
178
+ return torch.device(module._hf_hook.execution_device)
179
+ return self.device
180
+
181
+ def _encode_prompt(
182
+ self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
183
+ ):
184
+ r"""
185
+ Encodes the prompt into text encoder hidden states.
186
+
187
+ Args:
188
+ prompt (`str` or `list(int)`):
189
+ prompt to be encoded
190
+ device: (`torch.device`):
191
+ torch device
192
+ num_images_per_prompt (`int`):
193
+ number of images that should be generated per prompt
194
+ do_classifier_free_guidance (`bool`):
195
+ whether to use classifier free guidance or not
196
+ negative_prompt (`str` or `List[str]`):
197
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
198
+ if `guidance_scale` is less than `1`).
199
+ """
200
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
201
+
202
+ text_inputs = self.tokenizer(
203
+ prompt,
204
+ padding="max_length",
205
+ max_length=self.tokenizer.model_max_length,
206
+ truncation=True,
207
+ return_tensors="pt",
208
+ )
209
+ text_input_ids = text_inputs.input_ids
210
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
211
+
212
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
213
+ text_input_ids, untruncated_ids
214
+ ):
215
+ removed_text = self.tokenizer.batch_decode(
216
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
217
+ )
218
+ logger.warning(
219
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
220
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
221
+ )
222
+
223
+ if (
224
+ hasattr(self.text_encoder.config, "use_attention_mask")
225
+ and self.text_encoder.config.use_attention_mask
226
+ ):
227
+ attention_mask = text_inputs.attention_mask.to(device)
228
+ else:
229
+ attention_mask = None
230
+
231
+ text_embeddings = self.text_encoder(
232
+ text_input_ids.to(self.text_encoder.device), # FIXME 强制对齐device的位置
233
+ attention_mask=attention_mask,
234
+ )
235
+ text_embeddings = text_embeddings[0]
236
+
237
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
238
+ bs_embed, seq_len, _ = text_embeddings.shape
239
+ text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
240
+ text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
241
+
242
+ # get unconditional embeddings for classifier free guidance
243
+ if do_classifier_free_guidance:
244
+ uncond_tokens: List[str]
245
+ if negative_prompt is None:
246
+ uncond_tokens = [""] * batch_size
247
+ elif type(prompt) is not type(negative_prompt):
248
+ raise TypeError(
249
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
250
+ f" {type(prompt)}."
251
+ )
252
+ elif isinstance(negative_prompt, str):
253
+ uncond_tokens = [negative_prompt]
254
+ elif batch_size != len(negative_prompt):
255
+ raise ValueError(
256
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
257
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
258
+ " the batch size of `prompt`."
259
+ )
260
+ else:
261
+ uncond_tokens = negative_prompt
262
+
263
+ max_length = text_input_ids.shape[-1]
264
+ uncond_input = self.tokenizer(
265
+ uncond_tokens,
266
+ padding="max_length",
267
+ max_length=max_length,
268
+ truncation=True,
269
+ return_tensors="pt",
270
+ )
271
+
272
+ if (
273
+ hasattr(self.text_encoder.config, "use_attention_mask")
274
+ and self.text_encoder.config.use_attention_mask
275
+ ):
276
+ attention_mask = uncond_input.attention_mask.to(device)
277
+ else:
278
+ attention_mask = None
279
+
280
+ uncond_embeddings = self.text_encoder(
281
+ uncond_input.input_ids.to(self.text_encoder.device), # 同上,强制位置对齐。
282
+ attention_mask=attention_mask,
283
+ )
284
+ uncond_embeddings = uncond_embeddings[0]
285
+
286
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
287
+ seq_len = uncond_embeddings.shape[1]
288
+ uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
289
+ uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
290
+
291
+ # For classifier free guidance, we need to do two forward passes.
292
+ # Here we concatenate the unconditional and text embeddings into a single batch
293
+ # to avoid doing two forward passes
294
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
295
+
296
+ return text_embeddings
297
+
298
+ def decode_latents(self, latents):
299
+ b = latents.shape[0]
300
+ latents = 1 / 0.18215 * latents
301
+
302
+ is_video = len(latents.shape) == 5
303
+ if is_video:
304
+ latents = rearrange(latents, "b c f h w -> (b f) c h w")
305
+
306
+ image = self.vae.decode(latents).sample
307
+ image = (image / 2 + 0.5).clamp(0, 1)
308
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
309
+
310
+ image = image.cpu().float().numpy()
311
+ if is_video:
312
+ image = rearrange(image, "(b f) c h w -> b f h w c", b=b)
313
+ else:
314
+ image = rearrange(image, "b c h w -> b h w c")
315
+ return image
316
+
317
+ def prepare_extra_step_kwargs(self, generator, eta):
318
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
319
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
320
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
321
+ # and should be between [0, 1]
322
+
323
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
324
+ extra_step_kwargs = {}
325
+ if accepts_eta:
326
+ extra_step_kwargs["eta"] = eta
327
+
328
+ # check if the scheduler accepts generator
329
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
330
+ if accepts_generator:
331
+ extra_step_kwargs["generator"] = generator
332
+ return extra_step_kwargs
333
+
334
+ def check_inputs(self, prompt, height, width, callback_steps):
335
+ if not isinstance(prompt, str) and not isinstance(prompt, list):
336
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
337
+
338
+ if height % 8 != 0 or width % 8 != 0:
339
+ raise ValueError(
340
+ f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
341
+ )
342
+
343
+ if (callback_steps is None) or (
344
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
345
+ ):
346
+ raise ValueError(
347
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
348
+ f" {type(callback_steps)}."
349
+ )
350
+
351
+ def prepare_latents(
352
+ self,
353
+ batch_size,
354
+ num_channels_latents,
355
+ clip_length,
356
+ height,
357
+ width,
358
+ dtype,
359
+ device,
360
+ generator,
361
+ latents=None,
362
+ ):
363
+ if clip_length>0:
364
+ shape = (
365
+ batch_size,
366
+ num_channels_latents,
367
+ clip_length,
368
+ height // self.vae_scale_factor,
369
+ width // self.vae_scale_factor,
370
+ )
371
+ else:
372
+ shape = (
373
+ batch_size,
374
+ num_channels_latents,
375
+ height // self.vae_scale_factor,
376
+ width // self.vae_scale_factor,
377
+ )
378
+
379
+ if isinstance(generator, list) and len(generator) != batch_size:
380
+ raise ValueError(
381
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
382
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
383
+ )
384
+
385
+ if latents is None:
386
+ rand_device = "cpu" if device.type == "mps" else device
387
+
388
+ if isinstance(generator, list):
389
+ shape = (1,) + shape[1:]
390
+ latents = [
391
+ torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype)
392
+ for i in range(batch_size)
393
+ ]
394
+ latents = torch.cat(latents, dim=0).to(device)
395
+ else:
396
+ latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype).to(
397
+ device
398
+ )
399
+ else:
400
+ if latents.shape != shape:
401
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
402
+ latents = latents.to(device)
403
+
404
+ # scale the initial noise by the standard deviation required by the scheduler
405
+ latents = latents * self.scheduler.init_noise_sigma
406
+ return latents
407
+
408
+ @torch.no_grad()
409
+ def __call__(
410
+ self,
411
+ prompt: Union[str, List[str]],
412
+ height: Optional[int] = None,
413
+ width: Optional[int] = None,
414
+ fps_labels = None,
415
+ num_inference_steps: int = 50,
416
+ clip_length: int = 8,
417
+ guidance_scale: float = 7.5,
418
+ negative_prompt: Optional[Union[str, List[str]]] = None,
419
+ num_images_per_prompt: Optional[int] = 1,
420
+ eta: float = 0.0,
421
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
422
+ latents: Optional[torch.FloatTensor] = None,
423
+ output_type: Optional[str] = "pil",
424
+ return_dict: bool = True,
425
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
426
+ callback_steps: Optional[int] = 1,
427
+ ):
428
+ r"""
429
+ Function invoked when calling the pipeline for generation.
430
+
431
+ Args:
432
+ prompt (`str` or `List[str]`):
433
+ The prompt or prompts to guide the image generation.
434
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
435
+ The height in pixels of the generated image.
436
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
437
+ The width in pixels of the generated image.
438
+ num_inference_steps (`int`, *optional*, defaults to 50):
439
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
440
+ expense of slower inference.
441
+ guidance_scale (`float`, *optional*, defaults to 7.5):
442
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
443
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
444
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
445
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
446
+ usually at the expense of lower image quality.
447
+ negative_prompt (`str` or `List[str]`, *optional*):
448
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
449
+ if `guidance_scale` is less than `1`).
450
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
451
+ The number of images to generate per prompt.
452
+ eta (`float`, *optional*, defaults to 0.0):
453
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
454
+ [`schedulers.DDIMScheduler`], will be ignored for others.
455
+ generator (`torch.Generator`, *optional*):
456
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
457
+ to make generation deterministic.
458
+ latents (`torch.FloatTensor`, *optional*):
459
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
460
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
461
+ tensor will ge generated by sampling using the supplied random `generator`.
462
+ output_type (`str`, *optional*, defaults to `"pil"`):
463
+ The output format of the generate image. Choose between
464
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
465
+ return_dict (`bool`, *optional*, defaults to `True`):
466
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
467
+ plain tuple.
468
+ callback (`Callable`, *optional*):
469
+ A function that will be called every `callback_steps` steps during inference. The function will be
470
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
471
+ callback_steps (`int`, *optional*, defaults to 1):
472
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
473
+ called at every step.
474
+
475
+ Returns:
476
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
477
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
478
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
479
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
480
+ (nsfw) content, according to the `safety_checker`.
481
+ """
482
+ # 0. Default height and width to unet
483
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
484
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
485
+
486
+ # 1. Check inputs. Raise error if not correct
487
+ self.check_inputs(prompt, height, width, callback_steps)
488
+
489
+ # 2. Define call parameters
490
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
491
+ device = self._execution_device
492
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
493
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
494
+ # corresponds to doing no classifier free guidance.
495
+ do_classifier_free_guidance = guidance_scale > 1.0
496
+
497
+ # 3. Encode input prompt
498
+ text_embeddings = self._encode_prompt(
499
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
500
+ )
501
+
502
+ # 4. Prepare timesteps
503
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
504
+ timesteps = self.scheduler.timesteps
505
+
506
+ # 5. Prepare latent variables
507
+ num_channels_latents = self.unet.in_channels
508
+
509
+ latents = self.prepare_latents(
510
+ batch_size * num_images_per_prompt,
511
+ num_channels_latents,
512
+ clip_length,
513
+ height,
514
+ width,
515
+ text_embeddings.dtype,
516
+ device,
517
+ generator,
518
+ latents,
519
+ )
520
+ latents_dtype = latents.dtype
521
+
522
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
523
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
524
+
525
+ # 7. Denoising loop
526
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
527
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
528
+ for i, t in enumerate(timesteps):
529
+ # expand the latents if we are doing classifier free guidance
530
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
531
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
532
+ # print(latent_model_input.shape, )
533
+ # predict the noise residual
534
+ if fps_labels:
535
+ if isinstance(fps_labels, list):
536
+ fps_labels = torch.tensor(fps_labels).to(self.unet.device)
537
+ # 控制帧率
538
+ noise_pred = self.unet(
539
+ latent_model_input, t, encoder_hidden_states=text_embeddings, fps_labels=fps_labels,
540
+ ).sample.to(dtype=latents_dtype)
541
+ else:
542
+ noise_pred = self.unet(
543
+ latent_model_input, t, encoder_hidden_states=text_embeddings
544
+ ).sample.to(dtype=latents_dtype)
545
+
546
+ # perform guidance
547
+ if do_classifier_free_guidance:
548
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
549
+ noise_pred = noise_pred_uncond + guidance_scale * (
550
+ noise_pred_text - noise_pred_uncond
551
+ )
552
+
553
+ # compute the previous noisy sample x_t -> x_t-1
554
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
555
+
556
+ # call the callback, if provided
557
+ if i == len(timesteps) - 1 or (
558
+ (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
559
+ ):
560
+ progress_bar.update()
561
+ if callback is not None and i % callback_steps == 0:
562
+ callback(i, t, latents)
563
+
564
+ # 8. Post-processing
565
+ image = self.decode_latents(latents)
566
+ # image[:, 1:, :, :, :] = image[:, 1:, :, :, :] + image[:, 0:1, :, :, :] # 叠加残差
567
+
568
+ # 9. Run safety checker
569
+ has_nsfw_concept = None
570
+
571
+ # 10. Convert to PIL
572
+ if output_type == "pil":
573
+ image = self.numpy_to_pil(image)
574
+
575
+ if not return_dict:
576
+ return (image, has_nsfw_concept)
577
+
578
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
579
+
580
+ @staticmethod
581
+ def numpy_to_pil(images):
582
+ if len(images.shape)==5:
583
+ pil_images = []
584
+ for sequence in images:
585
+ pil_images.append(DiffusionPipeline.numpy_to_pil(sequence))
586
+ return pil_images
587
+ else:
588
+ return DiffusionPipeline.numpy_to_pil(images)
589
+
590
+
591
+ # 改写一下 model_index.json的保存内容, Unet是新定义的,直接保存会导致读取的时候出错~
592
+ def to_json_string(self) -> str:
593
+ from diffusers import __version__
594
+ import json
595
+ import numpy as np
596
+
597
+ config_dict = self._internal_dict if hasattr(self, "_internal_dict") else {}
598
+ config_dict["_class_name"] = self.__class__.__name__
599
+ config_dict["_diffusers_version"] = __version__
600
+
601
+ def to_json_saveable(value):
602
+ if isinstance(value, np.ndarray):
603
+ value = value.tolist()
604
+ return value
605
+
606
+ config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()}
607
+ if 'unet' in config_dict:
608
+ config_dict["unet"] = [
609
+ "diffusers",
610
+ "UNet2DConditionModel"
611
+ ]
612
+ if 'controlnet' in config_dict:
613
+ config_dict['controlnet'] = [
614
+ "diffusers",
615
+ "UNet2DConditionModel"
616
+ ]
617
+ # 覆盖
618
+ return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
model/video_diffusion/pipelines/pipeline_stable_diffusion_controlnet3d.py ADDED
@@ -0,0 +1,482 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Bytedance Ltd. and/or its affiliates
2
+
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+
16
+ from .pipeline_st_stable_diffusion import SpatioTemporalStableDiffusionPipeline
17
+ from typing import Callable, List, Optional, Union
18
+ from diffusers.schedulers import (
19
+ DDIMScheduler,
20
+ DPMSolverMultistepScheduler,
21
+ EulerAncestralDiscreteScheduler,
22
+ EulerDiscreteScheduler,
23
+ LMSDiscreteScheduler,
24
+ PNDMScheduler,
25
+ )
26
+ from transformers import DPTForDepthEstimation
27
+ from transformers import CLIPTextModel, CLIPTokenizer
28
+ from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
29
+ from diffusers import AutoencoderKL, DDIMScheduler, DDPMScheduler
30
+ import torch
31
+ from einops import rearrange, repeat
32
+ import decord
33
+ import cv2
34
+ import random
35
+ import numpy as np
36
+ from ..models.unet_3d_condition import UNetPseudo3DConditionModel
37
+ from ..models.controlnet3d import ControlNet3DModel
38
+
39
+
40
+ class Controlnet3DStableDiffusionPipeline(SpatioTemporalStableDiffusionPipeline):
41
+ def __init__(
42
+ self,
43
+ vae: AutoencoderKL,
44
+ text_encoder: CLIPTextModel,
45
+ tokenizer: CLIPTokenizer,
46
+ unet: UNetPseudo3DConditionModel,
47
+ controlnet: ControlNet3DModel,
48
+ scheduler: Union[
49
+ DDIMScheduler,
50
+ PNDMScheduler,
51
+ LMSDiscreteScheduler,
52
+ EulerDiscreteScheduler,
53
+ EulerAncestralDiscreteScheduler,
54
+ DPMSolverMultistepScheduler,
55
+ ],
56
+ annotator_model=None,
57
+
58
+ ):
59
+ super().__init__(vae, text_encoder, tokenizer, unet, scheduler)
60
+
61
+ self.annotator_model = annotator_model
62
+ self.controlnet = controlnet
63
+ self.unet = unet
64
+ self.vae = vae
65
+ self.tokenizer = tokenizer
66
+ self.text_encoder = text_encoder
67
+ self.scheduler = scheduler
68
+ self.register_modules(
69
+ vae=vae,
70
+ text_encoder=text_encoder,
71
+ tokenizer=tokenizer,
72
+ unet=unet,
73
+ controlnet=controlnet,
74
+ scheduler=scheduler,
75
+ )
76
+
77
+ @staticmethod
78
+ def get_frames_preprocess(data_path, num_frames=24, sampling_rate=1, begin_indice=0, return_np=False):
79
+ vr = decord.VideoReader(data_path,)
80
+ n_images = len(vr)
81
+ fps_vid = round(vr.get_avg_fps())
82
+ frame_indices = [begin_indice + i*sampling_rate for i in range(num_frames)] # 随机取n帧
83
+
84
+
85
+ while n_images <= frame_indices[-1]:
86
+ # 超过视频长度,采样率减小直至不超过。
87
+ sampling_rate -= 1
88
+ if sampling_rate == 0:
89
+ # NOTE 边界检查
90
+ return None, None
91
+ frame_indices = [i*sampling_rate for i in range(num_frames)]
92
+ frames = vr.get_batch(frame_indices).asnumpy()
93
+
94
+ if return_np:
95
+ return frames, fps_vid
96
+
97
+ frames = torch.from_numpy(frames).div(255) * 2 - 1
98
+ frames = rearrange(frames, "f h w c -> c f h w").unsqueeze(0)
99
+ return frames, fps_vid
100
+
101
+ @torch.no_grad()
102
+ def get_canny_edge_map(self, frames, ):
103
+ # (b f) c h w"
104
+ # from tensor to numpy
105
+ inputs = frames.cpu().numpy()
106
+ inputs = rearrange(inputs, 'f c h w -> f h w c')
107
+ # inputs from [-1, 1] to [0, 255]
108
+ inputs = (inputs + 1) * 127.5
109
+ inputs = inputs.astype(np.uint8)
110
+ lower_threshold = 100
111
+ higher_threshold = 200
112
+ edge_images = np.stack([cv2.Canny(inp, lower_threshold, higher_threshold) for inp in inputs])
113
+ # from numpy to tensors
114
+ edge_images = torch.from_numpy(edge_images).unsqueeze(1) # f, 1, h, w
115
+ edge_images = edge_images.div(255)*2 - 1
116
+ # print(torch.max(out_images), torch.min(out_images), out_images.dtype)
117
+ return edge_images.to(dtype= self.controlnet.dtype, device=self.controlnet.device)
118
+
119
+ @torch.no_grad()
120
+ def get_depth_map(self, frames, height, width, return_standard_norm=False ):
121
+ """
122
+ frames should be like: (f c h w), you may turn b f c h w -> (b f) c h w first
123
+ """
124
+ h,w = height, width
125
+ inputs = torch.nn.functional.interpolate(
126
+ frames,
127
+ size=(384, 384),
128
+ mode="bicubic",
129
+ antialias=True,
130
+ )
131
+ # 转类型和设备
132
+ inputs = inputs.to(dtype= self.annotator_model.dtype, device=self.annotator_model.device)
133
+
134
+ outputs = self.annotator_model(inputs)
135
+ predicted_depths = outputs.predicted_depth
136
+
137
+ # interpolate to original size
138
+ predictions = torch.nn.functional.interpolate(
139
+ predicted_depths.unsqueeze(1),
140
+ size=(h, w),
141
+ mode="bicubic",
142
+ )
143
+
144
+ # normalize output
145
+ if return_standard_norm:
146
+ depth_min = torch.amin(predictions, dim=[1, 2, 3], keepdim=True)
147
+ depth_max = torch.amax(predictions, dim=[1, 2, 3], keepdim=True)
148
+ predictions = 2.0 * (predictions - depth_min) / (depth_max - depth_min) - 1.0
149
+ else:
150
+ predictions -= torch.min(predictions)
151
+ predictions /= torch.max(predictions)
152
+
153
+ return predictions
154
+
155
+
156
+ @torch.no_grad()
157
+ def get_hed_map(self, frames,):
158
+ if isinstance(frames, torch.Tensor):
159
+ # 输入的就是 b c h w的tensor 范围是-1~1,需要转换为0~1
160
+ frames = (frames + 1) / 2
161
+ #rgb转bgr
162
+ bgr_frames = frames.clone()
163
+ bgr_frames[:, 0, :, :] = frames[:, 2, :, :]
164
+ bgr_frames[:, 2, :, :] = frames[:, 0, :, :]
165
+
166
+ edge = self.annotator_model(bgr_frames) # 范围也是0~1
167
+ return edge
168
+ else:
169
+ assert frames.ndim == 3
170
+ frames = frames[:, :, ::-1].copy()
171
+ with torch.no_grad():
172
+ image_hed = torch.from_numpy(frames).to(next(self.annotator_model.parameters()).device, dtype=next(self.annotator_model.parameters()).dtype )
173
+ image_hed = image_hed / 255.0
174
+ image_hed = rearrange(image_hed, 'h w c -> 1 c h w')
175
+ edge = self.annotator_model(image_hed)[0]
176
+ edge = (edge.cpu().numpy() * 255.0).clip(0, 255).astype(np.uint8)
177
+ return edge[0]
178
+
179
+ @torch.no_grad()
180
+ def get_pose_map(self, frames,):
181
+ if isinstance(frames, torch.Tensor):
182
+ # 输入的就是 b c h w的tensor 范围是-1~1,需要转换为0~1
183
+ frames = (frames + 1) / 2
184
+ np_frames = frames.cpu().numpy() * 255
185
+ np_frames = np.array(np_frames, dtype=np.uint8)
186
+ np_frames = rearrange(np_frames, 'f c h w-> f h w c')
187
+ poses = np.stack([self.annotator_model(inp) for inp in np_frames])
188
+ else:
189
+ poses = self.annotator_model(frames)
190
+ return poses
191
+
192
+ def get_timesteps(self, num_inference_steps, strength,):
193
+ # get the original timestep using init_timestep
194
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
195
+
196
+ t_start = max(num_inference_steps - init_timestep, 0)
197
+ timesteps = self.scheduler.timesteps[t_start:]
198
+
199
+ return timesteps, num_inference_steps - t_start
200
+
201
+ @torch.no_grad()
202
+ def __call__(
203
+ self,
204
+ prompt: Union[str, List[str]],
205
+ controlnet_hint = None,
206
+ fps_labels = None,
207
+ height: Optional[int] = None,
208
+ width: Optional[int] = None,
209
+ num_inference_steps: int = 50,
210
+ clip_length: int = 8, # NOTE clip_length和images的帧数一致。
211
+ guidance_scale: float = 7.5,
212
+ negative_prompt: Optional[Union[str, List[str]]] = None,
213
+ num_images_per_prompt: Optional[int] = 1,
214
+ eta: float = 0.0,
215
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
216
+ latents: Optional[torch.FloatTensor] = None,
217
+ output_type: Optional[str] = "pil",
218
+ return_dict: bool = True,
219
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
220
+ callback_steps: Optional[int] = 1,
221
+ cross_attention_kwargs = None,
222
+ video_scale: float = 0.0,
223
+ controlnet_conditioning_scale: float = 1.0,
224
+ fix_first_frame=True,
225
+ first_frame_output = None , # 也可以允许挑好图后传入。
226
+ first_frame_output_latent = None,
227
+ first_frame_control_hint = None, # 维持第一帧
228
+ add_first_frame_by_concat = False,
229
+ controlhint_in_uncond = False,
230
+ init_same_noise_per_frame=False,
231
+ init_noise_by_residual_thres=0.0,
232
+ images=None,
233
+ in_domain=False, # 是否调用视频模型生成图片
234
+ residual_control_steps=1,
235
+ first_frame_ddim_strength=1.0,
236
+ return_last_latent = False,
237
+ ):
238
+ '''
239
+ add origin video frames to get depth maps
240
+ '''
241
+
242
+ if fix_first_frame and first_frame_output is None and first_frame_output_latent is None:
243
+ first_frame_output = self.__call__(
244
+ prompt=prompt,
245
+ controlnet_hint=controlnet_hint[:,:,0,:,:] if not in_domain else controlnet_hint[:,:,0:1,:,:],
246
+ # b c f h w
247
+ num_inference_steps=20,
248
+ width=width,
249
+ height=height,
250
+ guidance_scale=guidance_scale,
251
+ num_images_per_prompt=1,
252
+ generator=generator,
253
+ fix_first_frame=False,
254
+ controlhint_in_uncond=controlhint_in_uncond,
255
+ ).images[0]
256
+
257
+
258
+ if first_frame_output is not None:
259
+ if isinstance(first_frame_output, list):
260
+ first_frame_output = first_frame_output[0]
261
+ first_frame_output = torch.from_numpy(np.array(first_frame_output)).div(255) * 2 - 1
262
+ first_frame_output = rearrange(first_frame_output, "h w c -> c h w").unsqueeze(0) # FIXME 目前不允许多个batch 先设置为1
263
+ first_frame_output = first_frame_output.to(dtype= self.vae.dtype, device=self.vae.device)
264
+
265
+ first_frame_output_latent = self.vae.encode(first_frame_output).latent_dist.sample()
266
+ first_frame_output_latent = first_frame_output_latent * 0.18215
267
+ # 0. Default height and width to unet
268
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
269
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
270
+
271
+ # 1. Check inputs. Raise error if not correct
272
+ self.check_inputs(prompt, height, width, callback_steps)
273
+
274
+ # 2. Define call parameters
275
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
276
+ device = self._execution_device
277
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
278
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
279
+ # corresponds to doing no classifier free guidance.
280
+ do_classifier_free_guidance = guidance_scale > 5.0
281
+
282
+ # 3. Encode input prompt
283
+ text_embeddings = self._encode_prompt(
284
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
285
+ )
286
+
287
+ # 4. Prepare timesteps
288
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
289
+ timesteps = self.scheduler.timesteps
290
+
291
+ # 5. Prepare latent variables
292
+ num_channels_latents = self.unet.in_channels
293
+ if controlnet_hint is not None:
294
+ if len(controlnet_hint.shape) == 5:
295
+ clip_length = controlnet_hint.shape[2]
296
+ else:
297
+ clip_length = 0
298
+
299
+ latents = self.prepare_latents(
300
+ batch_size * num_images_per_prompt,
301
+ num_channels_latents,
302
+ clip_length,
303
+ height,
304
+ width,
305
+ text_embeddings.dtype,
306
+ device,
307
+ generator,
308
+ latents,
309
+ )
310
+ latents_dtype = latents.dtype
311
+
312
+
313
+ if len(latents.shape) == 5 and init_same_noise_per_frame:
314
+ latents[:,:,1:,:,:] = latents[:,:,0:1,:,:]
315
+
316
+ if len(latents.shape) == 5 and init_noise_by_residual_thres > 0.0 and images is not None:
317
+
318
+ images = images.to(device=device, dtype=latents_dtype) # b c f h w
319
+ image_residual = torch.abs(images[:,:,1:,:,:] - images[:,:,:-1,:,:])
320
+ images = rearrange(images, "b c f h w -> (b f) c h w")
321
+
322
+ # norm residual
323
+ image_residual = image_residual / torch.max(image_residual)
324
+
325
+ image_residual = rearrange(image_residual, "b c f h w -> (b f) c h w")
326
+ image_residual = torch.nn.functional.interpolate(
327
+ image_residual,
328
+ size=(latents.shape[-2], latents.shape[-1]),
329
+ mode='bilinear')
330
+ image_residual = torch.mean(image_residual, dim=1)
331
+
332
+ image_residual_mask = (image_residual > init_noise_by_residual_thres).float()
333
+ image_residual_mask = repeat(image_residual_mask, '(b f) h w -> b f h w', b=batch_size)
334
+ image_residual_mask = repeat(image_residual_mask, 'b f h w -> b c f h w', c=latents.shape[1])
335
+
336
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
337
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
338
+
339
+ # 7. Denoising loop
340
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
341
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
342
+ if fix_first_frame:
343
+ if add_first_frame_by_concat:
344
+ if len(first_frame_output_latent.shape) == 4:
345
+ latents = torch.cat([first_frame_output_latent.unsqueeze(2), latents], dim=2)
346
+ else:
347
+ latents = torch.cat([first_frame_output_latent, latents], dim=2)
348
+ if first_frame_control_hint is not None:
349
+ controlnet_hint = torch.cat([first_frame_control_hint, controlnet_hint], dim=2)
350
+ else:
351
+ controlnet_hint = torch.cat([controlnet_hint[:,:,0:1 ,:,:], controlnet_hint], dim=2)
352
+
353
+ if controlhint_in_uncond:
354
+ controlnet_hint = torch.cat([controlnet_hint] * 2) if do_classifier_free_guidance else controlnet_hint
355
+ for i, t in enumerate(timesteps):
356
+ # expand the latents if we are doing classifier free guidance
357
+ if i<residual_control_steps and len(latents.shape) == 5 and init_noise_by_residual_thres > 0.0 and images is not None :
358
+ if first_frame_ddim_strength < 1.0 and i == 0 :
359
+ # NOTE DDIM to get the first noise
360
+ first_frame_output_latent_DDIM = first_frame_output_latent.clone()
361
+ full_noise_timestep, _ = self.get_timesteps(num_inference_steps, strength=first_frame_ddim_strength)
362
+ latent_timestep = full_noise_timestep[:1].repeat(batch_size * num_images_per_prompt)
363
+ first_frame_output_latent_DDIM = self.scheduler.add_noise(first_frame_output_latent_DDIM, latents[:,:,0,:,:], latent_timestep)
364
+ latents[:,:,0,:,:]=first_frame_output_latent_DDIM
365
+ begin_frame = 1
366
+ for n_frame in range(begin_frame, latents.shape[2]):
367
+ latents[:,:, n_frame, :, :] = \
368
+ (latents[:,:, n_frame, :, :] - latents[:,:, n_frame-1, :, :]) \
369
+ * image_residual_mask[:,:, n_frame-1, :, :] + \
370
+ latents[:,:, n_frame-1, :, :]
371
+ if fix_first_frame:
372
+ latents[:,:,0 ,:,:] = first_frame_output_latent
373
+
374
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
375
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
376
+ if controlnet_hint is not None:
377
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
378
+ latent_model_input,
379
+ t,
380
+ encoder_hidden_states=text_embeddings,
381
+ controlnet_cond=controlnet_hint,
382
+ return_dict=False,
383
+ )
384
+ down_block_res_samples = [
385
+ down_block_res_sample * controlnet_conditioning_scale
386
+ for down_block_res_sample in down_block_res_samples
387
+ ]
388
+ mid_block_res_sample *= controlnet_conditioning_scale
389
+
390
+ noise_pred = self.unet(
391
+ latent_model_input,
392
+ t,
393
+ encoder_hidden_states=text_embeddings,
394
+ cross_attention_kwargs=cross_attention_kwargs,
395
+ down_block_additional_residuals=down_block_res_samples,
396
+ mid_block_additional_residual=mid_block_res_sample,
397
+ ).sample.to(dtype=latents_dtype)
398
+ else:
399
+ # predict the noise residual
400
+ noise_pred = self.unet(
401
+ latent_model_input,
402
+ t,
403
+ encoder_hidden_states=text_embeddings,
404
+ ).sample.to(dtype=latents_dtype)
405
+
406
+ if video_scale > 0 and controlnet_hint is not None:
407
+ bsz = latents.shape[0]
408
+ f = latents.shape[2]
409
+ # 逐帧预测
410
+ latent_model_input_single_frame = rearrange(latent_model_input, 'b c f h w -> (b f) c h w')
411
+ text_embeddings_single_frame = torch.cat([text_embeddings] * f, dim=0)
412
+ control_maps_single_frame = rearrange(controlnet_hint, 'b c f h w -> (b f) c h w')
413
+ latent_model_input_single_frame = latent_model_input_single_frame.chunk(2, dim=0)[0]
414
+ text_embeddings_single_frame = text_embeddings_single_frame.chunk(2, dim=0)[0]
415
+ if controlhint_in_uncond:
416
+ control_maps_single_frame = control_maps_single_frame.chunk(2, dim=0)[0]
417
+
418
+ down_block_res_samples_single_frame, mid_block_res_sample_single_frame = self.controlnet(
419
+ latent_model_input_single_frame,
420
+ t,
421
+ encoder_hidden_states=text_embeddings_single_frame,
422
+ controlnet_cond=control_maps_single_frame,
423
+ return_dict=False,
424
+ )
425
+ down_block_res_samples_single_frame = [
426
+ down_block_res_sample_single_frame * controlnet_conditioning_scale
427
+ for down_block_res_sample_single_frame in down_block_res_samples_single_frame
428
+ ]
429
+ mid_block_res_sample_single_frame *= controlnet_conditioning_scale
430
+
431
+ noise_pred_single_frame_uncond = self.unet(
432
+ latent_model_input_single_frame,
433
+ t,
434
+ encoder_hidden_states = text_embeddings_single_frame,
435
+ down_block_additional_residuals=down_block_res_samples_single_frame,
436
+ mid_block_additional_residual=mid_block_res_sample_single_frame,
437
+ ).sample
438
+ noise_pred_single_frame_uncond = rearrange(noise_pred_single_frame_uncond, '(b f) c h w -> b c f h w', f=f)
439
+ # perform guidance
440
+ if do_classifier_free_guidance:
441
+ if video_scale > 0 and controlnet_hint is not None:
442
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
443
+ noise_pred = noise_pred_single_frame_uncond + video_scale * (
444
+ noise_pred_uncond - noise_pred_single_frame_uncond
445
+ ) + guidance_scale * (
446
+ noise_pred_text - noise_pred_uncond
447
+ )
448
+ else:
449
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
450
+ noise_pred = noise_pred_uncond + guidance_scale * (
451
+ noise_pred_text - noise_pred_uncond
452
+ )
453
+
454
+ # compute the previous noisy sample x_t -> x_t-1
455
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
456
+
457
+ # call the callback, if provided
458
+ if i == len(timesteps) - 1 or (
459
+ (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
460
+ ):
461
+ progress_bar.update()
462
+ if callback is not None and i % callback_steps == 0:
463
+ callback(i, t, latents)
464
+ # 8. Post-processing
465
+ image = self.decode_latents(latents)
466
+ if add_first_frame_by_concat:
467
+ image = image[:,1:,:,:,:]
468
+
469
+ # 9. Run safety checker
470
+ has_nsfw_concept = None
471
+ # 10. Convert to PIL
472
+ if output_type == "pil":
473
+ image = self.numpy_to_pil(image)
474
+
475
+ if not return_dict:
476
+ return (image, has_nsfw_concept)
477
+
478
+ if return_last_latent:
479
+ last_latent = latents[:,:,-1,:,:]
480
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept), last_latent
481
+ else:
482
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
requirements.txt ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ diffusers==0.14.0
2
+ transformers==4.27.3
3
+ accelerate==0.18.0
4
+ xformers==0.0.16
5
+ imageio==2.27.0
6
+ decord==0.6.0
7
+ opencv-python==4.7.0.72
8
+ einops==0.6.0