stargan-v2
stargan-v2 copied to clipboard
Training results cannot meet expectations
Hello, I used your settings to train on AFHQ and set the batchsize to 4. When switching to wild animals, the result was very poor. The FID index was still 60 at 80,000 steps.Can you help me answer my doubts? Thank you, I am very puzzled, because I want to conduct research based on your network, but the reproduction effect is not good, and I cannot proceed downward.
发现是因为batchsize的问题
Hi could you please share the hypeparameter you have used to avoid this problem (i.e. the batch size)? Thank you
Batch size must be 8,Only then can we get the result with very low FID value,The batchsize is 4, the effect will be very poor
So you only changed the batch size with respect to the default parameters ?
yes
When batch=4, training will become unstable.