File size: 17,079 Bytes
7ccc423
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
import inspect
from dataclasses import dataclass
from typing import Callable, List, Optional, Union

import numpy as np
import torch
from diffusers import DiffusionPipeline
from diffusers.image_processor import VaeImageProcessor
from diffusers.schedulers import (DDIMScheduler, DPMSolverMultistepScheduler,
                                  EulerAncestralDiscreteScheduler,
                                  EulerDiscreteScheduler, LMSDiscreteScheduler,
                                  PNDMScheduler)
from diffusers.utils import BaseOutput, is_accelerate_available
from diffusers.utils.torch_utils import randn_tensor
from einops import rearrange
from tqdm import tqdm
from transformers import CLIPImageProcessor

from src.models.mutual_self_attention import ReferenceAttentionControl


@dataclass
class Pose2VideoPipelineOutput(BaseOutput):
    videos: Union[torch.Tensor, np.ndarray]
    middle_results: Union[torch.Tensor, np.ndarray]


class Pose2VideoPipeline(DiffusionPipeline):
    _optional_components = []

    def __init__(
        self,
        vae,
        image_encoder,
        reference_unet,
        denoising_unet,
        pose_guider,
        scheduler: Union[
            DDIMScheduler,
            PNDMScheduler,
            LMSDiscreteScheduler,
            EulerDiscreteScheduler,
            EulerAncestralDiscreteScheduler,
            DPMSolverMultistepScheduler,
        ],
        image_proj_model=None,
        tokenizer=None,
        text_encoder=None,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            image_encoder=image_encoder,
            reference_unet=reference_unet,
            denoising_unet=denoising_unet,
            pose_guider=pose_guider,
            scheduler=scheduler,
            image_proj_model=image_proj_model,
            tokenizer=tokenizer,
            text_encoder=text_encoder,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.clip_image_processor = CLIPImageProcessor()
        self.ref_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
        )
        self.cond_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor,
            do_convert_rgb=True,
            do_normalize=False,
        )

    def enable_vae_slicing(self):
        self.vae.enable_slicing()

    def disable_vae_slicing(self):
        self.vae.disable_slicing()

    def enable_sequential_cpu_offload(self, gpu_id=0):
        if is_accelerate_available():
            from accelerate import cpu_offload
        else:
            raise ImportError("Please install accelerate via `pip install accelerate`")

        device = torch.device(f"cuda:{gpu_id}")

        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
            if cpu_offloaded_model is not None:
                cpu_offload(cpu_offloaded_model, device)

    @property
    def _execution_device(self):
        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
            return self.device
        for module in self.unet.modules():
            if (
                hasattr(module, "_hf_hook")
                and hasattr(module._hf_hook, "execution_device")
                and module._hf_hook.execution_device is not None
            ):
                return torch.device(module._hf_hook.execution_device)
        return self.device

    def decode_latents(self, latents):
        video_length = latents.shape[2]
        latents = 1 / 0.18215 * latents
        latents = rearrange(latents, "b c f h w -> (b f) c h w")
        # video = self.vae.decode(latents).sample
        video = []
        for frame_idx in tqdm(range(latents.shape[0])):
            video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
        video = torch.cat(video)
        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
        video = (video / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
        video = video.cpu().float().numpy()
        return video

    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(
            inspect.signature(self.scheduler.step).parameters.keys()
        )
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(
            inspect.signature(self.scheduler.step).parameters.keys()
        )
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        width,
        height,
        video_length,
        dtype,
        device,
        generator,
        latents=None,
    ):
        shape = (
            batch_size,
            num_channels_latents,
            video_length,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(
                shape, generator=generator, device=device, dtype=dtype
            )
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    def _encode_prompt(
        self,
        prompt,
        device,
        num_videos_per_prompt,
        do_classifier_free_guidance,
        negative_prompt,
    ):
        batch_size = len(prompt) if isinstance(prompt, list) else 1

        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(
            prompt, padding="longest", return_tensors="pt"
        ).input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
            text_input_ids, untruncated_ids
        ):
            removed_text = self.tokenizer.batch_decode(
                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
            )

        if (
            hasattr(self.text_encoder.config, "use_attention_mask")
            and self.text_encoder.config.use_attention_mask
        ):
            attention_mask = text_inputs.attention_mask.to(device)
        else:
            attention_mask = None

        text_embeddings = self.text_encoder(
            text_input_ids.to(device),
            attention_mask=attention_mask,
        )
        text_embeddings = text_embeddings[0]

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        bs_embed, seq_len, _ = text_embeddings.shape
        text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
        text_embeddings = text_embeddings.view(
            bs_embed * num_videos_per_prompt, seq_len, -1
        )

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

            max_length = text_input_ids.shape[-1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )

            if (
                hasattr(self.text_encoder.config, "use_attention_mask")
                and self.text_encoder.config.use_attention_mask
            ):
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            uncond_embeddings = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            uncond_embeddings = uncond_embeddings[0]

            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = uncond_embeddings.shape[1]
            uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
            uncond_embeddings = uncond_embeddings.view(
                batch_size * num_videos_per_prompt, seq_len, -1
            )

            # For classifier free guidance, we need to do two forward passes.
            # Here we concatenate the unconditional and text embeddings into a single batch
            # to avoid doing two forward passes
            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])

        return text_embeddings

    @torch.no_grad()
    def __call__(
        self,
        ref_image,
        pose_images,
        width,
        height,
        video_length,
        num_inference_steps,
        guidance_scale,
        num_images_per_prompt=1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        output_type: Optional[str] = "tensor",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: Optional[int] = 1,
        **kwargs,
    ):
        # Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        device = self._execution_device

        do_classifier_free_guidance = guidance_scale > 1.0

        # Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps

        batch_size = 1

        # Prepare clip image embeds
        clip_image = self.clip_image_processor.preprocess(
            ref_image, return_tensors="pt"
        ).pixel_values
        clip_image_embeds = self.image_encoder(
            clip_image.to(device, dtype=self.image_encoder.dtype)
        ).image_embeds
        encoder_hidden_states = clip_image_embeds.unsqueeze(1)
        uncond_encoder_hidden_states = torch.zeros_like(encoder_hidden_states)

        if do_classifier_free_guidance:
            encoder_hidden_states = torch.cat(
                [uncond_encoder_hidden_states, encoder_hidden_states], dim=0
            )
        reference_control_writer = ReferenceAttentionControl(
            self.reference_unet,
            do_classifier_free_guidance=do_classifier_free_guidance,
            mode="write",
            batch_size=batch_size,
            fusion_blocks="full",
        )
        reference_control_reader = ReferenceAttentionControl(
            self.denoising_unet,
            do_classifier_free_guidance=do_classifier_free_guidance,
            mode="read",
            batch_size=batch_size,
            fusion_blocks="full",
        )

        num_channels_latents = self.denoising_unet.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            width,
            height,
            video_length,
            clip_image_embeds.dtype,
            device,
            generator,
        )

        # Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # Prepare ref image latents
        ref_image_tensor = self.ref_image_processor.preprocess(
            ref_image, height=height, width=width
        )  # (bs, c, width, height)
        ref_image_tensor = ref_image_tensor.to(
            dtype=self.vae.dtype, device=self.vae.device
        )
        ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)

        # Prepare a list of pose condition images
        pose_cond_tensor_list = []
        for pose_image in pose_images:
            pose_cond_tensor = (
                torch.from_numpy(np.array(pose_image.resize((width, height)))) / 255.0
            )
            pose_cond_tensor = pose_cond_tensor.permute(2, 0, 1).unsqueeze(
                1
            )  # (c, 1, h, w)
            pose_cond_tensor_list.append(pose_cond_tensor)
        pose_cond_tensor = torch.cat(pose_cond_tensor_list, dim=1)  # (c, t, h, w)
        pose_cond_tensor = pose_cond_tensor.unsqueeze(0)
        pose_cond_tensor = pose_cond_tensor.to(
            device=device, dtype=self.pose_guider.dtype
        )
        pose_fea = self.pose_guider(pose_cond_tensor)
        pose_fea = (
            torch.cat([pose_fea] * 2) if do_classifier_free_guidance else pose_fea
        )

        # denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                # 1. Forward reference image
                if i == 0:
                    self.reference_unet(
                        ref_image_latents.repeat(
                            (2 if do_classifier_free_guidance else 1), 1, 1, 1
                        ),
                        torch.zeros_like(t),
                        # t,
                        encoder_hidden_states=encoder_hidden_states,
                        return_dict=False,
                    )
                    reference_control_reader.update(reference_control_writer)

                # 3.1 expand the latents if we are doing classifier free guidance
                latent_model_input = (
                    torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                )
                latent_model_input = self.scheduler.scale_model_input(
                    latent_model_input, t
                )

                noise_pred = self.denoising_unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=encoder_hidden_states,
                    pose_cond_fea=pose_fea,
                    return_dict=False,
                )[0]

                # perform guidance
                if do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + guidance_scale * (
                        noise_pred_text - noise_pred_uncond
                    )

                # call the callback, if provided
                if i == len(timesteps) - 1 or (
                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
                ):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
                        step_idx = i // getattr(self.scheduler, "order", 1)
                        callback(step_idx, t, latents)

            reference_control_reader.clear()
            reference_control_writer.clear()

        # Post-processing
        images = self.decode_latents(latents)  # (b, c, f, h, w)

        # Convert to tensor
        if output_type == "tensor":
            images = torch.from_numpy(images)

        if not return_dict:
            return images

        return Pose2VideoPipelineOutput(videos=images)