CycleGAN training tips

I’m going through CycleGAN example on fastai github repo.

Training itself was completed without errors but the results were not good.

This is the results of
learn.show_results(ds_type=DatasetType.Train, rows=2)

It looks results were not good even for training datasets.

Only thing I changed was batch size.
Since batch size 4 could not fit on my GPU which had 8GB memory, I changed it to 2.

Is this because batch size 2 is too small?
or because 100 epochs, which I didn’t change, is not enough?

Could you give me a hint?

Thanks in advance.

1 Like