models/eval.py

import os
import numpy
import torch
import random
from PIL import Image
import cv2
from diffusers import DDPMPipeline
from diffusers.utils import make_image_grid
import matplotlib.pyplot as plt
import numpy
from skimage.color import lab2rgb


def evaluate(condition_images, config, epoch, pipeline):
    # Sample some images from random noise (this is the backward diffusion process).
    # The default pipeline output type is `List[PIL.Image]`

    test_dir = os.path.join(config.output_dir, "samples")
    os.makedirs(test_dir, exist_ok=True)

    list_images = pipeline(
        condition_images=condition_images,
        batch_size=config.eval_batch_size,
        output_type="numpy",
        generator=torch.manual_seed(config.seed),
    ).images
    list_images = list_images * numpy.array([[[[100, 127, 127]]]])

    list_images = numpy.concatenate(
        (
            numpy.clip(list_images[..., :1], a_min=0, a_max=100),
            numpy.clip(list_images[..., 1:], a_min=-128, a_max=127),
        ),
        axis=3,
    )
    list_images = lab2rgb(list_images, channel_axis=3)
    list_images = (list_images * 255).astype(numpy.uint8)
    for i, image in enumerate(list_images):
        cv2.imwrite(f"{test_dir}/{epoch:04d}_{i:02d}_rgb.png", image)


def less_noisy_image(sample, timestep, betas, prev_timestep, model_output):
    alphas = 1.0 - betas
    alphas_cumprod = torch.cumprod(alphas, dim=0)
    alpha_prod_t = alphas_cumprod[timestep]
    # alpha_prod_t_prev = alphas_cumprod[prev_timestep] if prev_timestep >= 0 else 1.0
    alpha_prod_t_prev = (
        (torch.where(prev_timestep > 0, alphas_cumprod[prev_timestep], 1.0))
        .unsqueeze(-1)
        .unsqueeze(-1)
        .unsqueeze(-1)
    )

    beta_prod_t = 1 - alpha_prod_t
    beta_prod_t_prev = 1 - alpha_prod_t_prev

    alpha_prod_t = alpha_prod_t.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
    beta_prod_t = beta_prod_t.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)

    model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
    sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5)
    model_output_denom_coeff = alpha_prod_t.view(-1) * beta_prod_t_prev.view(-1) ** (
        0.5
    ) + (alpha_prod_t.view(-1) * beta_prod_t.view(-1) * alpha_prod_t_prev.view(-1)) ** (
        0.5
    )

    # print(sample_coeff.shape, sample.shape, alpha_prod_t_prev.shape, alpha_prod_t.shape, model_output.shape, model_output_denom_coeff.shape)
    prev_sample = sample_coeff * sample - (
        alpha_prod_t_prev - alpha_prod_t
    ) * model_output / model_output_denom_coeff.unsqueeze(-1).unsqueeze(-1).unsqueeze(
        -1
    )

    return prev_sample


def make_grid(images, rows, cols):
    w, h = images[0].size
    grid = Image.new("RGB", size=(cols * w, rows * h))
    for i, image in enumerate(images):
        grid.paste(image, box=(i % cols * w, i // cols * h))
    return grid