Closed simpleoier closed 6 years ago
@simpleoier yes, I guess you're right. I think I was mistaken by
in the last activation layer thereis a one-dimensional convolution layer with one filter of width one that does not downsample the hidden activations (1 × 1 convolution).
in the paper. But it seems clear from the code that there is a fully connected layer required before the output of discriminator.
Thanks for commenting; I'll put up a commit soon that fixes this :) (or you can open a pull request, if you don't mind)
Actually, in the paper
The latter (3) reduces the amount of parameters re- quired for the final classification neuron, which is fully con- nected to all hidden activations with a linear behavior.
I'll keep this open until the fix has been made
Fix made with commit 26f2a8e
. closing.
As a true or false indicator, shouldn't the output of discriminator be a scalar?