canonicalize/util.py

import os
import imageio
import numpy as np
from typing import Union
import cv2
import torch
import torchvision

from tqdm import tqdm
from einops import rearrange

def shifted_noise(betas, image_d=512, noise_d=256, shifted_noise=True):
    alphas = 1 - betas
    alphas_bar = torch.cumprod(alphas, dim=0)
    d = (image_d / noise_d) ** 2
    if shifted_noise:
        alphas_bar = alphas_bar / (d - (d - 1) * alphas_bar)
    alphas_bar_sqrt = torch.sqrt(alphas_bar)
    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
    # Shift so last timestep is zero.
    alphas_bar_sqrt -= alphas_bar_sqrt_T
    # Scale so first timestep is back to old value.
    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (
    alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

    # Convert alphas_bar_sqrt to betas
    alphas_bar = alphas_bar_sqrt ** 2
    alphas = alphas_bar[1:] / alphas_bar[:-1]
    alphas = torch.cat([alphas_bar[0:1], alphas])
    betas = 1 - alphas
    return betas

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, duration=1000/fps)

def save_imgs_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")
    for i, x in enumerate(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)
        os.makedirs(os.path.dirname(path), exist_ok=True)
        cv2.imwrite(os.path.join(path, f'view_{i}.png'), x[:,:,::-1])

def imgs_grid(videos: torch.Tensor, rescale=False, n_rows=4, fps=8):
    videos = rearrange(videos, "b c t h w -> t b c h w")
    image_list = []
    for i, x in enumerate(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)
        # image_list.append(x[:,:,::-1])
        image_list.append(x)
    return image_list

# 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.to(torch.float32), t, cond_embeddings.to(torch.float32), 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