New probability for Xs = Xt for AEI training

ocastan commented 4 years ago

The authors of the original paper have issued a new version in late June, and I noticed they corrected the training strategy of AEI and now Xs Xt are supposed to be different (and not identical anymore) 80% of the time So same,_prob=0.8 should be changed to 0.2 on line 56 of train_AEI.py I'm going to try this new value and I'll let you know

tangxuvis commented 4 years ago

hi， what is going on about your new experiments? any progresses？@ @ocastan

tangxuvis commented 4 years ago

i have implemented some perfomance, can we chat on email about more details? @ocastan
my email address: tangxuvis@163.com

ocastan commented 4 years ago

Hello My experiment is still running, each batch takes 3s. It has been running for about 80k iterations and the results are not very convincing so far with these new parameters. Moreover I"ve made numerous modifications:

gradient accumulation to simulate a batch of 24 samples with a single card
replaced batchnorm with groupnorm because of the previous
added gradient penalty (but it looks useless, it fires very rarely)
replaced mean of the tensor hinge with hinge of the tensor mean (not sure it was a good idea...)
... Can you tell a little bit more about your performance additions?

taotaonice / FaceShifter

New probability for Xs = Xt for AEI training #30