CycleGAN Notebook

Practice implementing CycleGAN in Jupyter Notebook.

Original paper

Official implementation

The image below shows the result of tensorflow model trained about 200 epochs, I think it is worse than the result of the original author.

Below is the result of pytorch model, trained about 135 epochs on horse2zebra.

Notebooks

Tensorflow 2.0 Notebook

Pytorch Notebook

Some Minor Modification

• Using random number within a range to replace label 0 and 1 while calculating the loss.

  For example, using the random number between 0 to 0.3 while computing the loss of fake images, and 0.7 to 1.2 for real images.This idea is from ganhacks. (Not sure whether it would be better for CycleGAN model or not)

Environments

• Python 3
• jupyter
• numpy
• matplotlib
• tensorflow 2.0
• pytorch 1.2.0
• tqdm

Get Started

1. Prepare Dataset

• Download the datasets in original author's paper.

  Check here.

  Take monet2photo for instance.

  Linux

  mkdir -p datasets/monet2photo/ wget https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/monet2photo.zip -O ./datasets/monet2photo.zip unzip ./datasets/monet2photo.zip -d ./datasets/ rm ./datasets/monet2photo.zip

  Windows

  Download the dataset from the link https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/monet2photo.zip, and unzip it in the datasets folder. The folder structure would be:

  .
  ├── ...
  ├── tf_cycle_gan.ipynb
  ├── datasets
  │   ├── monet2photo
  │         ├── testA
  │         ├── testB
  │         ├── trainA
  │         ├── trainB
  │
  └── ...
  

• Use own dataset(I have not try yet)

  It should be the same as the previous Windows part. First, you have to create a base folder in datasets, aldo have testA, testB, trainA, trainB folders in the base folder you just created. I think it is better to have close amout of images in trainA and trainB. Also as the original author says,

You should not expect our method to work on just any random combination of input and output datasets (e.g. cats<->keyboards). From our experiments, we >find it works better if two datasets share similar visual content. For example, landscape painting<->landscape photographs works much better than portrait >painting <-> landscape photographs. zebras<->horses achieves compelling results while cats<->dogs completely fails.

2. Download the pretrained checkpoints(Optional)

Download the pretrained checkpoints below, and unzip it in the checkpoints folder.

Monet2Photo

3. Modify the parameter in the notebook

Change the DATASET variable to the dataset name in the Hyper Parameters & Setting section.

4. Run the notebook

You can use the following command to check the status while training the model.

cd /to/path/of/this/project tensorboard --logdir logs

About 400~600 seconds per epoch with a NVIDIA 1080Ti GPU.

Known Issues

• Fail message in console: While training the models in tensorflow notebook, the console will appear this message meta_optimizer.cc:502 layout failed: Invalid argument: Size of values 0 does not match size of permutation 4., I don't know where is the error occur, this error message also appear in the Tensorflow Official CycleGAN Tutorial.

Acknowledgements

Thanks to:

• Official implementation
• Tensorflow Official CycleGAN Tutorial
• Pytorch DCGAN Tutorial

I reference some part of implementation from them.