search

cjbayesian / toyGANs

A toy model of Generative Adversarial Networks
24 stars 7 forks source link

regularisation cost #1

Open sitmo opened 7 years ago

sitmo commented 7 years ago

Hi Corey,

Thanks for your great blog post and sharing the code.

I was wondering about the weights in the the cost functions: shouldn't the l2_losses in the _generativeloss function have opposite signs? The way it is now increasing weights will reduce the cost in the generative_loss.

cjbayesian commented 7 years ago

Nice catch. I think you're right. I'm running it again with the sign change on the regularizers. Will push the change once it finishes.