Extract image to smaller images in dataloader

vision
1

How can I extract a image to smaller images in dataloader instead creating a new dataset that croped the image manually

    def __getitem__(self, index):
        img = Image.open(self.files[index % len(self.files)])
        img_lr = self.lr_transform(img)
        img_hr = self.hr_transform(img)

        return {"lr": img_lr, "hr": img_hr}

I want to extract the img(2048x2048) crop into four smaller images (1024x1024 each)

 __ __          __ __
|     |  =>    |__|__| 
|__ __|        |__|__|
2

If you don’t want to apply the cropping inside the Dataset.__getitem__, you could use it in the DataLoader loop:

for data, target in loader:
    # apply the transformation here

To create these four patches, you could either use unfold or torchvision.transforms.functional.five_crop and remove the center crop:

data = torch.randn(1, 1, 2048, 2048)

patches = data.unfold(2, 1024, 1024).unfold(3, 1024, 1024)
print(patches.shape)
> torch.Size([1, 1, 2, 2, 1024, 1024])
# reshape if needed

imgs = TF.five_crop(data, size=1024)
# remove center crop, imgs is a list, so you might want to `torch.stack` them
imgs = imgs[:4]