Pytorch-cyclegan-and-pix2pix: U-Net with different input sizes

[ShaniGam]

I'm new with U-Nets. What do I need to change in order to use it for 80x80 images?
Thanks.

[jiangwei221]

I tried maps dataset(600x600) and horse2zebra dataset(256x256), they both directly works using python train.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan (change maps to horse2zebra)
I think you do not need to do any changes to use it for 80x80 images.

[happytaoxiaoli]

@jiangwei221 do you know if it works when the training data is not eaqual in height and width?
maybe the pictures are 600 * 500 and not 600 * 600

[jiangwei221]

@happytaoxiaoli #49 this may help you

[happytaoxiaoli]

thanks very much @jiangwei221

[ShaniGam]

Unfortunately, it doesn't work. When I use unet_128 or unet_256 I get the error:
raise ValueError('Expected more than 1 value per channel when training, got input size {}'.format(size)) ValueError: Expected more than 1 value per channel when training, got input size [1L, 512L, 1L, 1L]
It works perfectly when I use one of the resnets.

[jiangwei221]

@ShaniGam Sorry, I tried cycle gan with resnet version, instead of the unet version.

[junyanz]

Unfortunately, UNet does not work with arbitrary input sizes. Try ResNet instead.

[ShaniGam]

Any ideas of what to change to make it work for 80x80 images?

[junyanz]

You can tweak the kernel size or padding of some conv layers. But you can resize your 80x80 to 128x128, and feed 128x128 to the network.

[ShaniGam]

OK, thanks!

[happsky]

Change '8' to '4' in net = UnetGenerator(input_nc, output_nc, 8, ngf, norm_layer=norm_layer, use_dropout=use_dropout) works for me when input 64*64 images.