use light mode G not work at all

[HUSTLX]

just like this, which D has a very small loss

[Frizy-up]

may need to train more steps

[HUSTLX]

may need to train more steps

already, it dose not work

[saturosfz]

I also met such case

[cfanyyx]

+1, I already trained more than 100000 steps but still got blank outputs.

[xiaosean]

+1, Using light mode gets similar results.

[cfanyyx]

It seems the input channel after adaptive_avg_pooling in light mode can't be set too small.

if self.light:
    FC = [nn.Linear(ngf * mult, ngf * mult, bias=False),  
          nn.ReLU(True),  
          nn.Linear(ngf * mult, ngf * mult, bias=False),  
          nn.ReLU(True)]

if I change the first ngf * mult to a larger one, then the output become normal.

[jayzhan211]

I try with this command python main.py --dataset selfie2anime --light True --resume True --save_freq 10000 --benchmark_flag True and the following is my result at 100,000 epoch. btw I continue the training start from 10000 (it is just the save network testing before).

B2A is terrible...