search

NVlabs / FUNIT

Translate images to unseen domains in the test time with few example images.
https://nvlabs.github.io/FUNIT/
Other
1.58k stars 238 forks source link

The training code does not work well... #40

Closed UdonDa closed 4 years ago

UdonDa commented 4 years ago

The result which is generated by the model is trained during 400 iterations.

I think it looks that the models fail to learn. I modify inplace=False in nn.ReLU and nn.LeakyReLU in blocks.py. following this issue https://github.com/NVlabs/FUNIT/issues/23

gen_train_00000400_02

Thanks.

orpatashnik commented 4 years ago

It takes a while for the model to generate some nice pictures. I trained with batch size of 16 pictures, and the pictures become reasonable after something like 20K iterations. It is possible to stabilize the training by reducing the learning rate, but then the final results look worse.

UdonDa commented 4 years ago

gen_train_00050500_03

I can get the clear samples! thanks!