bstriner / keras-adversarial

Keras Generative Adversarial Networks
MIT License
867 stars 231 forks source link

example_gan.py always generates same digit (1 in my case). #24

Open Phani-Nidadavolu opened 7 years ago

Phani-Nidadavolu commented 7 years ago

I am using Keras 2 and the latest keras-adverserial code that you wrote for keras-2. @bstriner Could you help me debug this?

mynameisvinn commented 7 years ago

theres no easy way to debug this issue.

in simple GANs, the discriminator has strong bias on what a "real" image looks like. as result, it teaches the generator to produce a single image, regardless of the initial input.

ive seen some research around minibatch discrimination, which makes sure features within various samples remain varied, and therefore avoids a single output.

kechan commented 6 years ago

I ran example_gan.py and I got mostly 1's and 7's. @mynameisvinn Could you elaborate a bit more if you know how to improve this? I don't get anything close to the picture show in this git.