search

neka-nat / image_completion_tf2

"Globally and Locally Consistent Image Completion" with Tensorflow2 Keras
31 stars 11 forks source link

data preprocess #15

Open LexieYang opened 4 years ago

LexieYang commented 4 years ago

Do we need to normalize the training images into the range of (-1,1) before feed them into the network?

neka-nat commented 4 years ago

Thank you for the suggestion. I couldn't find any such description in the paper, but I've seen it in other implementation, so I think your suggestion is correct.

Have you tried it?

dfaker commented 4 years ago

That would sound more like a tanh activation function rather than sigmoid, which is common for gans.