search

mit-han-lab / gan-compression

[CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs
Other
1.1k stars 150 forks source link

[Question] About SuperSeparableConv2d #113

Closed saduf closed 1 year ago

saduf commented 1 year ago

Hi, thanks for sharing this great work. In the regular Cyclegan Resnet block, both Conv2d use the same number channels, usually 64*4. In the SuperMobileResnetBlock, the first SuperSeparableConv2d uses the sampled number of channels from the resnet_configs channels-64-cycleGAN, and the second uses the input.size(1). I would expect both SuperSeparableConv2d to use the same number of channels as in the original generator.

I would appreciate your help understanding this, Thank you.

lmxyy commented 1 year ago

Hi, this is because the ResNet Block has a residual connection so the input channel number must match the output channel number. What we can do is just change the first convolution's output channel number. The others are all determined.

saduf commented 1 year ago

Thanks a lot for your response!