Conditional transforms for image resize

#1

I only want to resize images that are smaller than my desired input size. Is there way to reshape images that are smaller than a certain size and ignore all others?

#2

Using the Dataset class, you can lazily load your images in the __getitem__ method.
Here is a small example:

class TrainDataset(Dataset):
    def __init__(self, image_paths, targets):
        self.image_paths = image_paths
        self.targets = targets

    def __getitem__(self, index):
        image = Image.open(self.image_paths[index])
        y = self.target[index]

        # Resize your image here
        if image.size...:
            ....
        .... # Do other stuff here

        x = torch.from_numpy(image)
    
        return x, y

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

You can wrap your Dataset to a DataLoader for efficient loading, preprocessing, shuffling etc.

Hope this helps!

3 Likes
#3

For what it’s worth, the alternative to this is to subclass Resize. The advantage of this approach is that existing torchvision Dataset objects (like for CIFAR-10 or ImageNet) need not be modified - rather, a new transform can simply be used.

My implementation:

class ConditionalResize(transforms.Resize):
    """
    Resize transform but only if the input is smaller than the resize dims
    """

    @property
    def resize_height(self):
        return self.size if isinstance(self.size, int) else self.size[0]

    @property
    def resize_width(self):
        return self.size if isinstance(self.size, int) else self.size[1]

    def forward(self, img):
        """
        Args:
            img (PIL Image or Tensor): Image to be scaled.

        Returns:
            PIL Image or Tensor: Rescaled image.
        """
        r_h, r_w = self.resize_height, self.resize_width
        if isinstance(img, torch.Tensor):
            h, w = img.shape[1:]
        else:  # PIL Image
            w, h = img.size
        if w < r_w or h < r_h:
            return super().forward(img)
        else:
            return img
1 Like