Progressive growing of GANs

Inspired from NVIDIA’s 2017 paper “PROGRESSIVE GROWING OF GANS FOR IMPROVED QUALITY, STABILITY, AND VARIATION”: https://arxiv.org/pdf/1710.10196.pdf

The github page is here : https://github.com/tkarras/progressive_growing_of_gans

CelebA dataset

The celebA dataset can be downloaded on the project's home page. Here we’ll assume this was done and the pictures are all in a directory.

In[]:=

celebADirectory="/home/grondilu/Images/img_align_celeba";

We don’t import the images all in one go as that would take too much time. Instead, we create a list of symbolic file representations.

In[]:=

celebAFiles=File/@FileNames["*.jpg",celebADirectory];

Here is a random sample :

In[]:=

ImageCollage[Import/@RandomChoice[celebAFiles,9]]

Out[]=

These images are not square, but that’s ok, when we encode them we’ll squeeze them. In fact, the default NetEncoder does it for images anyway (notice the dimensions of the output):

In[]:=

Print@NetEncoder["Image"]

NetEncoder

Type:	Image
Output:	array(size: 3×128×128)



Also the default size for an image is 128x128, which is probably large enough for this experiment as I don’t have any NVIDIA GPU.

Network structure

In the paper they use a latent space of 512 dimensions.

In[]:=

latentSpaceSize=512;

The discriminator will have to scale the input image down. This can be done with the input NetEncoder :

In[]:=

NetEncoder[{"Image",{32,32}}][Import@RandomChoice@celebAFiles]//NetDecoder["Image"]

Out[]=

In[]:=

gan=NetGANOperator[{NetChain[{ReshapeLayer[{latentSpaceSize,1,1}],{ConvolutionLayer[latentSpaceSize,4,PaddingSize3],ParametricRampLayer[],ConvolutionLayer[latentSpaceSize,3,PaddingSize1],ParametricRampLayer[]},LinearLayer[{3,4,4}]},"Input"->latentSpaceSize,"Output"->NetDecoder["Image"]],NetChain[{{ConvolutionLayer[latentSpaceSize,3,PaddingSize1],ParametricRampLayer[],ConvolutionLayer[latentSpaceSize,3,PaddingSize1],ParametricRampLayer[]},LinearLayer[{}]},"Input"->NetEncoder[{"Image",{4,4}}]]}]

Out[]=

NetGANOperator