the preprocessor for detection, only resizing

[mohamedsaeed8223]

I have some cases where the images' dimensions are far from being like a square, their height are far larger than their width or the other way around

wouldn't it be better for those cases if the Document object or the preprocessor had the option to have them padded the image to the size we set to the model to take in?

maybe check the original width to original height ratio and if its too high or too low opt to padding the image

[felixdittrich92]

Hey @mohamedsaeed8223 👋 ,

That's already the case and enabled by default.
From docs:

preserve_aspect_ratio: if you want to preserve the aspect ratio of your documents while resizing before sending them to the model.

symmetric_pad: if you choose to preserve the aspect ratio, it will pad the image symmetrically and not from the bottom-right.

Both are set to True as default values.

Best,
Felix :)

[mohamedsaeed8223]

Thanks for replying @felixdittrich92 !💌

I found our key words preserve_aspect_ratio , symmetric_pad in the detection predictor class here

but they seem to be attributes of the torchvision.transforms.Resize object, and got no traces of them being set to True

although they could be passed as the extra **kwargs in the preprocessor constructor here

but looking in the torchvision docs if the Resize object could have those , and It seems like it cant. and those are completely different augments with their separate classes.

It would be highly appreciated if you could refer me to the code that can set those parameters to True, or how I could set them to be true while still using the preprocessor and predictor APIs.

Edit: it looks like doctr's custom Resize is used instead of torchvision's , but the values of preserve_aspect_ratio and symmetric_pad values are set to false by default here

its clear now how I could set them to be true:)

Thanks

[felixdittrich92]

doctr/doctr/transforms/modules/pytorch.py

Line 20 in 72bed41

class Resize(T.Resize):

About the args for the detection_predictor that's already updated to make it clearer for users on main branch:

doctr/doctr/models/detection/zoo.py

Line 78 in 72bed41

def detection_predictor(

Hope this helps :)