search

shaoanlu / faceswap-GAN

A denoising autoencoder + adversarial losses and attention mechanisms for face swapping.
3.39k stars 841 forks source link

Q: What's the reasoning behind using PixelShuffler over Conv2DTranspose? #40

Open mmittek opened 6 years ago

mmittek commented 6 years ago

The upscale block uses pixel shuffler after 4x convolution. I understand that this is a neat way of increasing the number of coefficients and then nicely reshaping everything to bring it one resolution step up, but why this and not Conv2DTranspose?

def upscale_ps(filters, use_norm=True): def block(x): x = Conv2D(filters*4, kernel_size=3, use_bias=False, kernel_initializer=RandomNormal(0, 0.02), padding='same' )(x) x = LeakyReLU(0.1)(x) x = PixelShuffler()(x) return x return block

shaoanlu commented 6 years ago

I remember I’ve tried convtranspose in v1 model, but I dont remember why I switched back to sub-pix (pixelshuffler)... Maybe it is because pixelshuffler is faster. I think it will not be much different in terms of final output quality though.