src/utils.py

import os

import torch
import matplotlib.pyplot as plt
from PIL import Image
from torch.utils.data import Dataset
from torchvision import transforms


class ImageDataset(Dataset):
    def __init__(self, dir, transform=None) -> None:
        self.dir = dir
        self.transform = transform
        self.file_list = sorted(os.listdir(self.dir))
    
    def __len__(self):
        return len(self.file_list)
    
    def __getitem__(self, idx):
        image_name = self.file_list[idx]
        image_path = os.path.join(self.dir, image_name)

        grayscale_image = Image.open(image_path).convert('L')
        colorized_image = Image.open(image_path).convert('RGB')

        if self.transform:
            grayscale_image = self.transform(grayscale_image)
            colorized_image = self.transform(colorized_image)
        
        return grayscale_image, colorized_image
    

def show_image(image_tensor):
    try:
        if len(image_tensor) == 1:
            plt.imshow(image_tensor[0], cmap="gray")
        else:
            plt.imshow(image_tensor.numpy().transpose(1, 2, 0))
    except Exception as e:
        print(f"Exception when showing image: {e}")


# To be able to calculate MSE loss in case output tensor has different shape from target tensor
def adjust_output_shape(output_tensor, target_tensor):
    adjusted_tensor = torch.nn.functional.interpolate(output_tensor, size=target_tensor.shape[2:], mode="bilinear", align_corners=False)
    return adjusted_tensor


def pil_to_torch(pil_image):
    transform = transforms.ToTensor()
    return transform(pil_image).unsqueeze(0)

def torch_to_pil(torch_image):
    transform = transforms.ToPILImage()
    return  transform(torch_image.squeeze(0))