Generative Adversarial Networks(GANs)

1. Vanilla GAN
2. ClusterGAN

Vanilla GAN

Model Structure

Final Generator Structure

• A MLP with 2 hidden layers of hidden_size=1024

• A LeakyReLU of slope=0.2 is used for each layer for activation

• nn.Sequential(
      nn.Linear(self.n_z, self.hidden_size),
      nn.LeakyReLU(0.2),
  
      nn.Linear(self.hidden_size, self.hidden_size),
      nn.LeakyReLU(0.2),
  
      nn.Linear(self.hidden_size, self.hidden_size),
      nn.LeakyReLU(0.2),
  
      nn.Linear(self.hidden_size, 784),
      nn.Tanh(),
  )

Final Discriminator Structure

• A MLP with 2 hidden layers of hidden_size=1024

• A LeakyReLU of slope=0.2 is used for each layer for activation

• A Dropout of rate=0.3 is used for each layer

• nn.Sequential(
      nn.Linear(self.n_input, self.hidden_size),
      nn.LeakyReLU(0.2),
      nn.Dropout(0.3),
  
      nn.Linear(self.hidden_size, self.hidden_size),
      nn.LeakyReLU(0.2),
      nn.Dropout(0.3),
  
      nn.Linear(self.hidden_size, self.hidden_size),
      nn.LeakyReLU(0.2),
      nn.Dropout(0.3),
  
      nn.Linear(self.hidden_size, 1),
      nn.Sigmoid(),
  )

I tried several settings of Generator and Discriminator for 200 epochs to determine the final structures, including activations, Dropout rate, and # of hidden layers.

Training Losses

Trained for 500 epochs on a GPU

Generated Results During Training

Trained for 500 epochs on a GPU

ClusterGAN

Model Structure

Adding an Encoder: X - > Z

Final Generator Structure

nn.Sequential(
    nn.Linear(z_dim + num_class, 1024),
    nn.BatchNorm1d(1024),
    nn.LeakyReLU(0.2),
    nn.Linear(1024, prod),
    nn.BatchNorm1d(prod),
    nn.LeakyReLU(0.2),

    Reshape((128, 7, 7)),

    nn.ConvTranspose2d(128, 64, 4, stride=2, padding=1, bias=True),
    nn.BatchNorm2d(64),
    nn.LeakyReLU(0.2),

    nn.ConvTranspose2d(64, 1, 4, stride=2, padding=1, bias=True),
    nn.Sigmoid()
)

Final Encoder Structure

nn.Sequential(
    nn.Conv2d(1, 64, 4, stride=2, bias=True),
    nn.LeakyReLU(0.2),
    nn.Conv2d(64, 128, 4, stride=2, bias=True),
    nn.LeakyReLU(0.2),

    Reshape((prod,)),

    torch.nn.Linear(prod, 1024),
    nn.LeakyReLU(0.2),
    torch.nn.Linear(1024, z_dim + num_class)
)

Final Discriminator Structure

nn.Sequential(
    nn.Conv2d(1, 64, 4, stride=2, bias=True),
    nn.LeakyReLU(0.2),
    nn.Conv2d(64, 128, 4, stride=2, bias=True),
    nn.LeakyReLU(0.2),

    Reshape((prod,)),

    nn.Linear(prod, 1024),
    nn.LeakyReLU(0.2),
    nn.Linear(1024, 1),
    nn.Sigmoid(),
)

Training Losses

Trained for 500 epochs on a GPU

Generator&Encoder seem to be not strong enough

Generated Results During Training

Trained for 500 epochs on a GPU

For Cluster GAN, I make the noise for display ordered by 10 classes

Reference

• [GPAM+14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014.
• [MALK19] Sudipto Mukherjee, Himanshu Asnani, Eugene Lin, and Sreeram Kannan. Clustergan: Latent space clustering in generative adversarial networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 4610–4617, 2019.
• https://github.com/eriklindernoren/PyTorch-GAN from PapersWithCode