Skip to content

togheppi/CycleGAN

Repository files navigation

CycleGAN

PyTorch implementation of CycleGAN

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
( FF8900 : Discriminator A / AE0000 : Discriminator B
CD52A7 : Generator A / 2156C9 : Generator B
58BEE4 : Cycle loss A / 319F92 : 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
( FF8900 : Discriminator A / AE0000 : Discriminator B
CD52A7 : Generator A / 2156C9 : Generator B
58BEE4 : Cycle loss A / 319F92 : 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