backward MMD

[adrienchaton]

Hello Schelotto

I have a short question ... why

total_loss = recon_loss - mmd_loss total_loss.backward()

and not + mmd_loss as both should be descended in the optimization step ?

Little remark about the RBF kernel, pytorch was throwing me an error because of in place modification before backward.

maybe worth changing res1 += torch.exp(-C * dists_y) by res1 = res1+torch.exp(-C * dists_y) ?

Lastly, is the previously posted issue on IMQ kernel solved ?

Thanks ! Adrien