search

martinarjovsky / WassersteinGAN

BSD 3-Clause "New" or "Revised" License
3.21k stars 725 forks source link

Questions about loss curve #31

Closed ghost closed 7 years ago

ghost commented 7 years ago

Hi! @martinarjovsky

I have some questions about the loss curve. When I don't use the median filter operation, the loss curve is still oscillates, which describes the training is unstable, am I right??

Please help me! Thank you!

martinarjovsky commented 7 years ago

Not really, that's because it's a training curve (as opposed to a validation set curve)! Since the minibatches are tiny, different minibatches will have different losses, hence the oscilation. If you were to plot this values over a big validation set (say ~ 10K examples), then the oscilations would mostly dissapear. This is expensive though, so we just put it through a simple filtering :).