Train fcdd with rectangular image

Question

Train fcdd with rectangular image

thy789 opened this issue a year ago · comments

Hi liznerski,

I'm trying to train a custom dataset. I wonder if I could use rectangular images to train fcdd. I tried using a rectangular size and faced an assertion error in this line: https://github.com/liznerski/fcdd/blob/97400156cd06604a2fa0189bee3dec5f65966f02/python/fcdd/models/bases.py#L162C1-L162C65

Do I simply remove the "in_shape[1] == in_shape[2]" portion to continue training with rectangular image or is there any other changes to make?

Thanks.

liznerski · Answer 1 · Mon Oct 23 2023 19:00:41 GMT+0800 (China Standard Time)

Hi!

In principle FCDD can handle this. However, the code expects the images to be square. I think the only issue is that the ReceptiveNet just tracks the height of the input. The receptive upsampling would thus not produce images of correct resolution. As long as you still use square convolutions, I think the rest should be fine.

Cheers.