Inspired from https://github.com/Ahmkel/Keras-Project-Template
CycleGAN is implemented as a sample model.
- tensorpack powered dataloader
- supports save/load models and optimizers
- supports multi-gpu training
- backups source code and config used to train
- stops training if encounter nan loss
- collages losses to a single graph on tensorboard
- telegram notification on train start/end
python >= 3.6
pip install -r requirements.txt
sh ./download_dataset.sh horse2zebra
edit configs/cyclegan.yml file
python train.py --config configs/cyclegan.yml
python train.py --config configs/cyclegan.yml --chkpt 123
tensorboard --logdir /path/to/experiment/dir/tensorboard/
Add telegram token and chat_id to environment variables as TELEGRAM_TOKEN and TELEGRAM_CHAT_ID