Closed EiffL closed 4 years ago
Actually, I've fixed a small issue in my implementation of the loss function (I hadn't seen this trick using the softplus at the discriminator output before), it seems much better behaved now, but we'll see in the morning if it works
I think that's the issue. This spectral normalization only works with this loss function or the other clipping version proposed in the paper, which is weird :p.
Chun-Liang
On Thu, Jan 10, 2019 at 5:12 AM Francois Lanusse notifications@github.com wrote:
Actually, I've fixed a small issue in my implementation of the loss function (I hadn't seen this trick using the softmax at the discriminator output before), it seems much better behaved now, but we'll see in the morning if it works
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/McWilliamsCenter/deep_galaxy_models/issues/10#issuecomment-453041757, or mute the thread https://github.com/notifications/unsubscribe-auth/AUsZHkeKEIb-Clv5PNWMC8cDE_dYxFzkks5vBxHwgaJpZM4Z5HVI .
Hi @chunliangli, I've tried to merge the SN code you pushed to the rest of the package, it runs as of now, but it's a bit misbehaving from times to times, not sure I did a good job at porting it over. But I did manage to train a generator that worked pretty well with the script you pushed. I've added 2 scripts, one that trains on COSMOS images, one that (in principle) trains on SDSS images (see https://github.com/McWilliamsCenter/deep_galaxy_models/blob/master/scripts/prepare_kaggle_data.py for a script that turns the SDSS data downloaded from the Kaggle website into a nice tfrecord that is then ingested by
train_sndcgan_sdss.py
). Both of them are living in branch u/EiffL/merge_GAN : https://github.com/McWilliamsCenter/deep_galaxy_models/tree/u/EiffL/merge_GANI'll try to give it another go tomorrow, but in case you get a chance to have a look, that would useful.