3_ControlNet/build_dataset.py

import os
import random
import math
import numpy as np
from PIL import Image
from torch.utils.data import Dataset
from torchvision import transforms


class CustomCocoDataset(Dataset):
    def __init__(self, img_folder, img_size=512, hint_size=448):
        self.img_folder = img_folder
        self.img_size = img_size
        self.hint_size = hint_size
        self.ids = [os.path.splitext(f)[0] for f in os.listdir(img_folder) if f.endswith(('.jpg', '.jpeg', '.png'))]

    def __len__(self):
        return len(self.ids)

    def __getitem__(self, index):
        img_id = self.ids[index]
        img_path = os.path.join(self.img_folder, img_id + '.png')
        image = Image.open(img_path).convert('RGB')

        # Perform a random crop using the custom random_crop_arr function
        cropped_image = random_crop_arr(image, self.img_size, min_crop_frac=0.8, max_crop_frac=1.0)

        # Convert cropped image back to PIL for further processing
        cropped_image = Image.fromarray(cropped_image)

        # Resize to different resolutions
        jpg_image = transforms.functional.to_tensor(cropped_image)
        hint_image = transforms.functional.resize(cropped_image, (self.hint_size, self.hint_size), interpolation=transforms.InterpolationMode.BICUBIC)
        hint_image = transforms.functional.to_tensor(hint_image)

        # Set captions to an empty string
        prompt = ""

        return dict(jpg=jpg_image, txt=prompt, hint=hint_image)

def random_crop_arr(pil_image, image_size, min_crop_frac=0.8, max_crop_frac=1.0):
    min_smaller_dim_size = math.ceil(image_size / max_crop_frac)
    max_smaller_dim_size = math.ceil(image_size / min_crop_frac)
    smaller_dim_size = random.randrange(min_smaller_dim_size, max_smaller_dim_size + 1)

    # We are not on a new enough PIL to support the reducing_gap
    # argument, which uses BOX downsampling at powers of two first.
    # Thus, we do it by hand to improve downsample quality.
    while min(*pil_image.size) >= 2 * smaller_dim_size:
        pil_image = pil_image.resize(
            tuple(x // 2 for x in pil_image.size), resample=Image.BOX
        )

    scale = smaller_dim_size / min(*pil_image.size)
    pil_image = pil_image.resize(
        tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC
    )

    arr = np.array(pil_image)
    crop_y = random.randrange(arr.shape[0] - image_size + 1)
    crop_x = random.randrange(arr.shape[1] - image_size + 1)
    return arr[crop_y : crop_y + image_size, crop_x : crop_x + image_size]


if __name__ == "__main__":
    dataset = CustomCocoDataset("/home/t2vg-a100-G4-1/projects/dataset/LSDIR_raw/images/train")
    print(len(dataset))
    print(dataset[0])

    from torch.utils.data import DataLoader
    dataloader = DataLoader(
        dataset, batch_size=4, num_workers=2,
        pin_memory=True, drop_last=True)
    
    # 从 DataLoader 中取出一个批次
    batch = next(iter(dataloader))

    # 提取批次中的 jpg_image 和 hint_image
    jpg_images = batch['jpg']
    hint_images = batch['hint']
    prompts = batch['txt']

    # 打印提示语
    print(f"Prompt: {prompts}")

    # 可视化并保存第一个batch的图像
    import matplotlib.pyplot as plt

    for i in range(len(jpg_images)):
        plt.figure(figsize=(10, 5))
    
        plt.subplot(1, 2, 1)
        plt.title(f"JPG Image {i+1} (512x512)")
        plt.imshow(jpg_images[i].permute(1, 2, 0))  # 转换维度以便imshow使用

        plt.subplot(1, 2, 2)
        plt.title(f"Hint Image {i+1} (448x448)")
        plt.imshow(hint_images[i].permute(1, 2, 0))  # 转换维度以便imshow使用

        # 保存图像到文件
        plt.savefig(f'output_image_{i+1}.png')

        # 关闭当前图像，释放内存
        plt.close()