search

mit-han-lab / gan-compression

[CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs
Other
1.1k stars 148 forks source link

SuperConv output channels #55

Closed 0x366 closed 3 years ago

0x366 commented 3 years ago

Hi! Nice paper!

I ran into issue, while trying to run train_supernet_lite code on my setup. SuperConv2d outputs variable number of channels (and doesn't depend on out_channels arg) While norm_layer's channels are always constant. So when I run SuperMobileResnetGenerator I always get running_mean error because of different number of channels in Batchnorm and SuperConv2d. How is this supposed to work ? 

from models.modules.resnet_architecture.super_mobile_resnet_generator import SuperMobileResnetGenerator
 from configs.resnet_configs import get_configs as get_super_configs
 import torch
 net =  SuperMobileResnetGenerator(4,3, ngf=64, n_blocks=9)
 x = torch.rand(1,4,256,256)
 super_config = get_super_configs("channels-64-pix2pix")
 net.configs = super_config.sample()
 y = net(x)

It raises error because Batchnorm has 64 channels, while output from Superconv has 48 channels (RuntimeError: running_mean should contain 48 elements not 64)

0x366 commented 3 years ago

My bad. You need to use InsanceNorm2d to make it work