how to train fully connected layers?

NVlabs / FUNIT

Translate images to unseen domains in the test time with few example images.

https://nvlabs.github.io/FUNIT/

Other

1.57k stars 238 forks source link

how to train fully connected layers? #8

Open hujinsen opened 5 years ago

hujinsen commented 5 years ago

When calculating affine parameters, Z_y is input into three full-connection layers, and then the mean and standard deviation are output. Why do we do this? How do full-connection layers train?

layumi commented 5 years ago

Hi @hujinsen , I think the MLP part works as a simple mapping function before we use them as the std and mean parameters of the adaptive instance normalization layer.

It is similar to the module in some other works, e.g., StyleGAN and MUNIT.