Generative_Adversarial_Networks

Various GAN networks implemented with tensorflow/pytorch

• Pytorch Implementation: Yoo Jaehoon

• Tensorflow Implementation: juiceleee

Result of InfoGAN-pytorch

Requirements

opencv-python==3.3.0.10
Pillow==4.3.0
torch==0.4.0
tensorflow-gpu==1.7.0
tensorflow-tensorboard==0.4.0rc3

Pytorch Network Architecture

Architectures are Different with Papers

• Vanilla_GAN
  • D
    • Conv2d(1, 32, 3, 1), Relu
    • Conv2d(32, 64, 3, 1), Relu
    • FC(28x28x64, 625)
    • FC(625, 1)
    • dropout(0.5)
    • Sigmoid
  • G
    • FC(100, 256), Relu, BatchNorm
    • FC(256, 512), Relu, BatchNorm
    • FC(512, 28x28)
    • Sigmoid
• DCGAN
  • D
    • Conv2d(1, 32, 3, 1), Relu
    • Conv2d(32, 64, 3, 1), Relu
    • FC(28x28x64, 625)
    • FC(625, 1)
    • dropout(0.5)
    • Sigmoid
  • G
    • FC(100, 7x7x16)
    • ConvTranspose2d(16, 4, 2, 2), Relu, BatchNorm
    • ConvTranspose2d(4, 1, 2, 2)
    • Sigmoid
• InfoGAN
  • D_front
    • Conv2d(1, 32, 3, 1), LeakyRelu(0.1)
    • Conv2d(32, 64, 3, 1), LeakyRelu(0.1)
    • dropout(0.3)
    • FC(28x28x64, 625), LeakyRelu(0.1)
    • for D
      • D_front
      • FC(625, 1)
      • dropout(0.5)
      • Sigmoid
    • for Q_class
      • D_front
      • FC(625, 10), Softmax
    • for Q_cont
      • D_front
      • FC(625, 2), Sigmoid
  • G
    • FC(62+12, 7x7x16)
    • ConvTranspose2d(16, 4, 2, 2), Relu, BatchNorm
    • ConvTranspose2d(4, 1, 2, 2)
    • Sigmoid

Reference

• Vanilla_GAN : [https://arxiv.org/abs/1406.2661]

• DCGAN : [https://arxiv.org/abs/1511.06434]

• InfoGAN : [https://arxiv.org/abs/1606.03657]