keras-acgan
keras-acgan copied to clipboard
lukedeo
Decreasing learning rate, to avoid divergence.
Hmm, seems reasonable - I'm going to push a few changes to this pretty soon (i.e., actual transposed convs, etc.) but seems harmless enough. What was your hardware setup that led to black images for the original learning rate as suggested in the DCGAN paper?
I used gtx 1080 with 357 driver version. Back end was tensorflow, however, I think I tried Theano as well.
On Feb 16, 2017, at 9:04, Luke de Oliveira [email protected] wrote:
Hmm, seems reasonable - I'm going to push a few changes to this pretty soon (i.e., actual transposed convs, etc.) but seems harmless enough. What was your hardware setup that led to black images for the original learning rate as suggested in the DCGAN paper?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.
Observed the same black generated images when running on GeForce GTX 1080 with 367.57 driver. It was so with both of TensorFlow (0.12.1) and Theano (0.9.0b1) backends.
