Look at your test results
middlexu opened this issue · comments
middlexu commented
It doesn't look good. The pictures are all dark.
why it doesn't work?
Alec Graves commented
GANs often have stability issues, and this little demo setup is no exception! I had to play with quite a few parameters to get even the results that I did. I have found that if the learning rate is too high, the models completely fail, and you can get a black image problem.
It may also help to modify how the images are pre-processed. I believe I just left the mean of the input data at whatever it naturally is. If you modify the input data's distribution to [-1, 1], this could also help the models overcome the dark image problem:
images = images.astype(keras.backend.floatx())
# divide by the largest uint8 value, setting the distribution to [0, 2]
images *= 2/255
# shift the distribution to [-1, 1]
images -= 1
and to reset the data to uint8 for viewing:
# shift the distribution to [0, 2]
images += 1
# shift the distribution back to [0, 255]
images *= 255/2 # restore to uint8
images[images < 0] = 0
images[images > 255] = 255
images = np.uint8(images)