cycleGAN inference pictures differ from images generated in training

[ptrblck]

You could create a suitable input e.g. by using x = torch.randn(your_shape) and save it via torch.save.
To compare the activations you could use forward hooks as described here.