Pytorch-cyclegan-and-pix2pix: Instance norm during evaluation

[HectorAnadon]

Hello,

I notice that during training, when sampling the images the model is not in eval mode https://pytorch.org/docs/stable/nn.html#torch.nn.Module.eval. Then, as track_running_stats is True in InstanceNorm, the model should behave differently when it is training or evaluating. I might have missed something.

[naoto0804]

I'm also curious about this. When I visit this repository a year ago, there was no track_running_stats in PyTorch. Does track_running_stats affect the performance?

[junyanz]

This is a good point. This is a new flag that we haven't tested. Maybe we should set the default value as Fasle so that we can match the original code. @SsnL @taesung89

[taesungp]

I agree. Let's set it to false. @HectorAnadon thanks for pointing this out! We really appreciate it.

[junyanz]

I updated the flag.