What is the difference between preprocess scale_width_and_crop and resize_and_crop?

[bclaw524]

The description seems to be the same where both resize the image to load_size and then crop the image to crop_size.

The only thing I seem to notice is that scale_width_and_crop seems to complete an epoch a bit faster than resize_and_crop.

(my images are 1024x1024 and the load size is 512)

[Jiwno]

in the file "./data/base_dataset.py", you can easily find the part to precess the image you imported.
If you use resize_and_crop option, width and height(new_h & new_w) are eqaul to a load_size.

but if you use scale_width option, height of load image change refer to the ration of the original image.

As far as I know, resize_and_crop option loads the images as square, same width and height, but scale_width_and_crop option loads the image as rectangular with aspect ration of the original data.

[noooob-ody]

i found it maybe a mistake in the source code
def __scale_width(img, target_size, crop_size, method=transforms.InterpolationMode.BICUBIC):
method = __transforms2pil_resize(method)
ow, oh = img.size
if ow == target_size and oh >= crop_size:
return img
w = target_size
h = int(max(target_size * oh / ow, crop_size))
return img.resize((w, h), method)
usually the width is bigger than height, so it still result in distortion, change it into def __scale_width(img, target_size, crop_size, method=transforms.InterpolationMode.BICUBIC):
method = __transforms2pil_resize(method)
ow, oh = img.size
if ow == target_size and oh >= crop_size:
return img
h = target_size
w = int(max(target_size * ow / oh, crop_size))
return img.resize((w, h), method)
and it keeps the aspect ratio