Pix2Pix Code - Generator and Discriminator questions about keras-gan HOT 4 CLOSED

Hi,

1. If you print the shape of the output of the last downsample layer in the generator you will see that the output has the shape (?, 2, 2, 512). You can do this by adding print (d7.shape) in Pix2Pix.build_generator. By setting padding='same' padding will be applied to preserve the spatial resolution of the input given that strides=1. If you use the same setting for padding and set strides=2 padding will be applied in the same way as in the previous case, but as stride is set to 2 the spatial resolution of the output is downsampled by a factor of two.

2. As the input image is propogated through four downsample layers the output shape of the discriminator will be (batch_size, img_height // 2**4, img_width // 2**4, 1) = (batch_size, 256 // 2**4, 256 // 2**4, 1) = (batch_size, 16, 16, 1).

from keras-gan.

Thanks for these answers!!

1. All these years I've thought "same" was ensuring spatial resolution regardless of stride by automatically padding the the input feature maps with zeros...

But I just checked deeper and you are absolutely right "same" is ONLY maintains spatial when stride=1. Since I always use stride equals one, i was confused! Thanks for clarifying this...

2. Your answer to number one answers this question of mine as well...

Thanks so much!!!!!!!

from keras-gan.

Awesome. No problem!

from keras-gan.

I think I answered my own question about the Patch GAN idea by the way. Is the idea that rather than just letting D output a single number for the entire image as in the traditional sense, you instead force smaller sub patches of the image to be "believable" by D and as a result produce higher quality images from G. This seems to make a lot of sense to me. What is your recommendation for tuning the ouptut spatial resolution of D, is higher better or worse?

It looks like you make use of MSE losses as well is this due to the insights from Least Squares GANs paper (which I thought was fantastic by the way)? I agree with your choice of loss functions just wondering what your rationale is...

THIS CODE IS AMAZING and thanks to your efforts I think I have a 90%+ understanding of what is going with pix2pix.

Concerning this Mode Collapse we always observe is there anything explicit in the code that I am missing that explicitly combats that? Ive heard of this term "mean distance between batch" but I dont know if that actually works well in practice.

Im closing this now but I would love your thoughts still!!

THANKS AGAIN!!!!!!!!!!!!!

EDIT: Just noticed you closed this already.

from keras-gan.