Pytorch implementation of our MCMI image-to-image translation method. Our proposed mutual information constraints produce higher quality images and learn more semantically-relevant mappings.
To install all requirements:
pip install -r requirements.txt
- Python 3.7
- cat2dog dataset is already available in ./datasets
- cat2dog: 871 cat (birman) images, 1364 dog (husky, samoyed) images crawled and cropped from Google Images.
To train the model(s) in the paper, run this command from src folder:
cd src
sh scripts/train_cyclegan.sh
This trains a MCMI-CycleGAN model on Cat2Dog dataset and saves the model every 50 epochs. You can train it on other datasets by modifying the dataroot in train_cyclegan.sh
- part of the training code is modified from official implementation of CycleGAN
- Evaluation code is in the evaluation_cyclegan folder
- Move trained MCMI-CycleGAN model to checkpoints/chat2dog_cyclegan folder for evaluation
- To evaluate the trained model for FID1, FID2 and LPIPS scores as described in the paper, run this command from evaluation_cyclegan folder:
cd evaluation_cyclegan
sh test_drit.sh
The evaluation code will look for model trained at epoch 200 under /evaluation_cyclegan/checkpoints/cat2dog_cyclegan folder. You need to change "all_files" in line 70 of text.py to evaluate at your own model epoch.
- It can take some time to download and load inception and alexnet weights
- We use the correct pytorch FID implementation from here: https://github.com/mseitzer/pytorch-fid
Our pre-trained model achieves the following performance on cat2dog dataset. You can download it from here: https://drive.google.com/file/d/1AXBwxl8MaWEiSl3etNX5f_yMYIKdQiTK/view?usp=sharing
| FID1 | FID2 | LPIPS |
|---|---|---|
| 62.85 | 28.09 | 0.22 |