This repository contains CycleGAN implementation code. The repository has GAN and CycleGAN architectures along with their test results.
The repository doesn't contain CycleGAN models and tested datasets due to their sizes. It has a simple GAN model in the models folder.
Open the jupyter file code/v3_GAN.ipynb to see the GAN structure tested on the MNIST dataset. This model is saved and it can be used without training as shown in the jupyter notebook.
In addition to GAN and CycleGAN scripts, the repository also has processing code for the day2night dataset, test results for a pretrained pix2pix model on the day2night dataset, and the evaluation code for calculating MSE and SSIM scores.
Test results from the datasets can be found in code/v5_cycleGAN/Pictures/Test/.
- Clone the repository.
- Create folder data
- Put your dataset under folder data/dataset_name.
- It should have subfolders as testA, testB, trainA, trainB with corresponding images.
- Create folders named as:
- code/v5_cycleGAN/Pictures/Test/dataset_name/testA/
- code/v5_cycleGAN/Pictures/Test/dataset_name/testB/
- code/v5_cycleGAN/Pictures/Training/dataset_name/
- models/v5_CycleGAN/
- Modify dataset variable in code/v5_cycleGAN/train.py, code/v5_cycleGAN/test.py, and code/v5_cycleGAN/plot.py to dataset_name.
- Run code/v5_cycleGAN/train.py.
- It saves losses in code/v5_cycleGAN/loss_arr/ and logs in code/v5_cycleGAN/loss_log/.
- It saves random generated images after each epoch in code/v5_cycleGAN/Pictures/Training/dataset_name/.
- It saves model in models/v5_CycleGAN/.
- Run code/v5_cycleGAN/test.py.
- It saves generated images in code/v5_cycleGAN/Pictures/Test/dataset_name/.
- Run code/v5_cycleGAN/plot.py for a training loss plot.