chrisdonahue / sdgan

Official implementation of "Semantically Decomposing the Latent Spaces of Generative Adversarial Networks"
MIT License
95 stars 19 forks source link

Sdgan-began goes into mode collapse #1

Open ltt-gddxz opened 6 years ago

ltt-gddxz commented 6 years ago

I run sdgan-began, at about 95000th step, it goes into mode collapse, and can not get satisfactory generated images as that in the paper.

I did not modify the code except that transforming the code from python2 to python3 using 2to3 command.

ltt-gddxz commented 6 years ago

Before mode collapse(at 61916th step), the result image seems promissing.

image

Anyway, thanks for this work.

chrisdonahue commented 6 years ago

Update 04/08/18

I'm reopening this issue because it is an ongoing work in progress for this repo. Our SD-BEGAN result from our paper was based on a fork of BEGAN-tensorflow. To reproduce our paper results, please see this zip file which contains the fork and a README.

I am working on a reimplementation in the began branch which should be faster and easier to use. However, as you pointed out, it currently suffers from mode collapse. I will continue to work to identify the issue, and contributions are welcome.

ltt-gddxz commented 6 years ago

Thanks, the zip file code works well, looking forward to the improvement.