G acc: 0.0000

[JSen]

Elapsed time in update: 1.213263
Iteration: 00065911/00100000
D acc: 0.9995 G acc: 0.0000
Elapsed time in update: 1.134419
Iteration: 00065912/00100000
D acc: 0.9999 G acc: 0.0000
Elapsed time in update: 1.168156
Iteration: 00065913/00100000
D acc: 1.0000 G acc: 0.0000
Elapsed time in update: 1.190268
Iteration: 00065914/00100000
D acc: 0.9999 G acc: 0.0000
Elapsed time in update: 1.129889
Iteration: 00065915/00100000
D acc: 1.0000 G acc: 0.0000
Elapsed time in update: 1.128156
Iteration: 00065916/00100000
D acc: 0.9999 G acc: 0.0000
Elapsed time in update: 1.135906
Iteration: 00065917/00100000
D acc: 1.0000 G acc: 0.0000
Elapsed time in update: 1.108964
Iteration: 00065918/00100000

G acc is 0 , is this normal?

[MarStarck]

same problem. my D performs too good to train G.
I recommend you to:

1. decrease D's lr

2. run K iters on G and 1 iter on D.

After doing these, my results got better but still not satisfactory.
And the original code use same lr and iter, I wonder why it can succeed

[Johnson yue]

@MarStarck Hi, In GAN theory ， the D is measure real distribution and fake distribution. So, it not make sense about run K iters on G and 1 iter on D. Contrary ， you should run K iters on D and 1 iter on G

[MarStarck]

@Johnson-yue yes, you are right. But in my experiment, this operation indeed improves performance when G acc is 0.
It's really strange because in theory calc_grad2 term can ensure no gradient vanishing problem in my understanding.

[Johnson yue]

yes， but just in theory or it depends on your dataset，I want to know does it improves generator real data??