CycleGAN: Horse to Zebra

Current update status

• Implement basic form of CycleGAN
• Train discriminators if their loss is bigger than criterion
• Replay buffer for training discriminator
• Different GAN loss (ex. WGAN-GP)

Model

Generator: ResNet (9 Residual Blocks)

Encoder(Downsampling) → Transformer(Residual Block x9) → Decoder(Upsampling)

Encoder:

Conv #1: ReflectionPad(3) → Conv(in_channels, 64, 7) → InstanceNorm → ReLU
Conv #2: Conv(64, 128, 3, 2, 1) → InstanceNorm → ReLU
Conv #3: Conv(128, 256, 3, 2, 1) → InstanceNorm → ReLU

Residual Block:

ReflectionPad(1) → Conv(256, 256, 3) → InstanceNorm → ReLU
→ ReflectionPad(1) → Conv(256, 256, 3) → InstanceNorm → Add

Decoder:

Conv #1: ConvT(256, 128, 3, 2, 1, out_padding=1) → InstanceNorm → ReLU
Conv #2: ConvT(128, 64, 3, 2, 1, out_padding=1) → InstanceNorm → ReLU
Conv #3: ReflectionPad(3) → Conv(64, out_channels, 7) → Tanh

Discriminator: PatchGAN (30x30)

Conv #1: Conv(in_channels, 64, 4, 2, 1) → LeakyReLU(0.2)
Conv #2: Conv(64, 128, 4, 2, 1) → BatchNorm → LeakyReLU(0.2)
Conv #3: Conv(128, 256, 4, 2, 1) → BatchNorm → LeakyReLU(0.2)
Conv #4: Conv(256, 512, 4, 1, 1) → BatchNorm → LeakyReLU(0.2)
Conv #5: Conv(512, 1, 4, 1, 1) → Sigmoid

Gan Loss: Mean Squared Error (= LSGAN)

---

Output images during training

---

Loss vs. epoch

---

Test