CycleGAN

PyTorch implementation of CycleGAN

• Dataset can be downloaded from here.
• Loss values are plotted using Tensorboard in PyTorch.

horse2zebra dataset

• Image size: 256x256
• Number of training images: 1,334 for horse images, 1,067 for zebra images
• Number of test images: 120 for horse images, 140 for zebra images

Results

• Adam optimizer is used. Learning rate = 0.0002, batch size = 1, # of epochs = 200:
• 6 resnet blocks used for Generator.

GAN losses ( : Discriminator A / : Discriminator B : Generator A / : Generator B : Cycle loss A / : Cycle loss B )	Generated images (Input / Generated / Reconstructed)

• Generated images using test data

  Horse to Zebra 1st column: Input / 2nd column: Generated / 3rd column: Reconstructed
  Zebra to Horse 1st column: Input / 2nd column: Generated / 3rd column: Reconstructed

apple2orange dataset

• Image size: 256x256
• Number of training images: 1,019 for apple images, 995 for orange images
• Number of test images: 266 for apple images, 248 for orange images

Results

• Adam optimizer is used. Learning rate = 0.0002, batch size = 1, # of epochs = 200:
• 9 resnet blocks used for Generator.

GAN losses ( : Discriminator A / : Discriminator B : Generator A / : Generator B : Cycle loss A / : Cycle loss B )	Generated images (Input / Generated / Reconstructed)

• Generated images using test data

  Apple to Orange 1st column: Input / 2nd column: Generated / 3rd column: Reconstructed
  Orange to Apple 1st column: Input / 2nd column: Generated / 3rd column: Reconstructed

References

1. https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
2. https://github.com/znxlwm/pytorch-CycleGAN
3. https://hardikbansal.github.io/CycleGANBlog