File size: 4,588 Bytes
8502051
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import imageio
import numpy as np
from typing import Union
import decord
decord.bridge.set_bridge('torch')
import torch
import torchvision
import PIL
from typing import List
from tqdm import tqdm
from einops import rearrange

from controlnet_aux import CannyDetector

def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=4, fps=8):
    videos = rearrange(videos, "b c t h w -> t b c h w")
    outputs = []
    for x in videos:
        x = torchvision.utils.make_grid(x, nrow=n_rows)
        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
        if rescale:
            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
        x = (x * 255).numpy().astype(np.uint8)
        outputs.append(x)

    os.makedirs(os.path.dirname(path), exist_ok=True)
    imageio.mimsave(path, outputs, fps=fps)

def save_videos_grid_pil(videos: List[PIL.Image.Image], path: str, rescale=False, n_rows=4, fps=8):
    videos = rearrange(videos, "b c t h w -> t b c h w")
    outputs = []
    for x in videos:
        x = torchvision.utils.make_grid(x, nrow=n_rows)
        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
        if rescale:
            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
        x = (x * 255).numpy().astype(np.uint8)
        outputs.append(x)

    os.makedirs(os.path.dirname(path), exist_ok=True)
    imageio.mimsave(path, outputs, fps=fps)

def read_video(video_path, video_length, width=512, height=512, frame_rate=2):
    vr = decord.VideoReader(video_path, width=width, height=height)
    sample_index = list(range(0, len(vr), frame_rate))[:video_length]
    video = vr.get_batch(sample_index)
    video = rearrange(video, "f h w c -> f c h w")
    video = (video / 127.5 - 1.0)
    return video


def get_annotation(video, annotator):
    t2i_transform = torchvision.transforms.ToPILImage()
    annotation = []
    for frame in video:
        pil_frame = t2i_transform(frame)
        if isinstance(annotator, CannyDetector):
            annotation.append(annotator(pil_frame, low_threshold=100, high_threshold=200))
        else:
            annotation.append(annotator(pil_frame))
    return annotation

# DDIM Inversion
@torch.no_grad()
def init_prompt(prompt, pipeline):
    uncond_input = pipeline.tokenizer(
        [""], padding="max_length", max_length=pipeline.tokenizer.model_max_length,
        return_tensors="pt"
    )
    uncond_embeddings = pipeline.text_encoder(uncond_input.input_ids.to(pipeline.device))[0]
    text_input = pipeline.tokenizer(
        [prompt],
        padding="max_length",
        max_length=pipeline.tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_embeddings = pipeline.text_encoder(text_input.input_ids.to(pipeline.device))[0]
    context = torch.cat([uncond_embeddings, text_embeddings])

    return context


def next_step(model_output: Union[torch.FloatTensor, np.ndarray], timestep: int,
              sample: Union[torch.FloatTensor, np.ndarray], ddim_scheduler):
    timestep, next_timestep = min(
        timestep - ddim_scheduler.config.num_train_timesteps // ddim_scheduler.num_inference_steps, 999), timestep
    alpha_prod_t = ddim_scheduler.alphas_cumprod[timestep] if timestep >= 0 else ddim_scheduler.final_alpha_cumprod
    alpha_prod_t_next = ddim_scheduler.alphas_cumprod[next_timestep]
    beta_prod_t = 1 - alpha_prod_t
    next_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
    next_sample_direction = (1 - alpha_prod_t_next) ** 0.5 * model_output
    next_sample = alpha_prod_t_next ** 0.5 * next_original_sample + next_sample_direction
    return next_sample


def get_noise_pred_single(latents, t, context, unet):
    noise_pred = unet(latents, t, encoder_hidden_states=context)["sample"]
    return noise_pred


@torch.no_grad()
def ddim_loop(pipeline, ddim_scheduler, latent, num_inv_steps, prompt):
    context = init_prompt(prompt, pipeline)
    uncond_embeddings, cond_embeddings = context.chunk(2)
    all_latent = [latent]
    latent = latent.clone().detach()
    for i in tqdm(range(num_inv_steps)):
        t = ddim_scheduler.timesteps[len(ddim_scheduler.timesteps) - i - 1]
        noise_pred = get_noise_pred_single(latent, t, cond_embeddings, pipeline.unet)
        latent = next_step(noise_pred, t, latent, ddim_scheduler)
        all_latent.append(latent)
    return all_latent


@torch.no_grad()
def ddim_inversion(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt=""):
    ddim_latents = ddim_loop(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt)
    return ddim_latents