VAEGAN training failing

Question

VAEGAN training failing

vishwakftw opened this issue 7 years ago · comments

Vishwak Srinivasan commented 7 years ago

The VAEGAN training fails after about 30 iterations.

Vishwak Srinivasan · Answer 1 · Tue Jul 04 2017 11:52:47 GMT+0800 (China Standard Time)

The training is stabilized. The L^th layer loss is decreasing, but very slowly. The images are incomprehensible.

Vishwak Srinivasan · Answer 2 · Sat Jul 22 2017 20:18:56 GMT+0800 (China Standard Time)

I think the issue has been fixed. The problem was with the loss function specified in the paper. L_{gan} is supposed to be of the opposite sign since we are minimizing it for the discriminator, as specified in the algorithm. The code was written verbatim from the text, hence the problem.

Vishwak Srinivasan · Answer 3 · Tue Aug 01 2017 08:03:50 GMT+0800 (China Standard Time)

However, after proposing the above fix, the image that are learned are incomprehensible. Below is the graph of the variation of all losses with the iterations.

Optimizer used is Adam with learning rate: 0.0005 and (beta₁, beta₂) = (0.5, 0.9)

Any help would be appreciated.