search

artcg / BEGAN

Boundary Equibilibrium Generative Adversarial Networks Implementation in Tensorflow
MIT License
208 stars 43 forks source link

bug? test time: only constant gray images #13

Closed AlexBinder closed 4 years ago

AlexBinder commented 7 years ago

Hi Artur,

there were some minor fixed necessary

add-epochs=0 -> --add-epochs=0 --train=False --> --train=0 also save the images as png, not jpg but then still the images were constant grey, as an plt.imshow(im_to_save) plt.show() could verify

values were all around 0.49...0.50, normalizing them to 0...1 gives only color clouds

AlexBinder commented 7 years ago

tensorflow 1.2.0 rc1

anastasiadk commented 7 years ago

I have the same issue as well. I changed a couple of things (removed non-linearities, subtracted 0.5), as suggested in this issue related to carpedm20/BEGAN-tensorflow, but still it doesn't converge and output at test time is all gray images, as AlexBinder stated above. Any idea about why this happens? Thanks in advance.

ChesterAiGo commented 6 years ago

I got the same problem...testing using the original CelebA dataset now..have you guys solved the problem? It might come from the usage of our custom dataset i'd say.

EDIT: Tested on a subset of CelebA (2600 images), for 10 epochs (I did not expect any outstanding results by using 10 epochs, but at least it should not be all grey...) with all default configs and it still outputs grey images...weird

artcg commented 4 years ago

Closing this issue which turned out to be a tensorflow version issue