The architecture of the generator network of InfoGAN is illustrated as follows. The reproduction of results from the original paper is rather tricky to handle. Therefore, we present a model architecture based on this GitHub repository, https://github.com/eriklindernoren/PyTorch-GAN/blob/master/implementations/infogan/infogan.py:
Generator architecture of InfoGAN
The generator network of InfoGAN consists of 4 hidden layers. The first hidden layer transforms the input vector with a length of 74 (62+10+2) into a length of 8,192 (128*8*8), which is then directly turned into tensor with the dimensionality of 128*8*8. ...
