Implementation of Memory Efficient Video GAN.
Install requirements.
pip install requirements.txt
Place ProGAN repo within MeVGAN repo.
git clone https://github.com/facebookresearch/pytorch_GAN_zoo progan
Apply patch to ProGAN repo. This will stop visdom session from running in the background, set batch size to 8 for ProGAN training, and add functionality to load ProGAN from checkpoint.
cd progan
git apply ../progan.patch
Copy model files from progan/models to MeVGAN/models to recreate original directory structure used to save ProGAN checkpoint. This is necessary to load model weigths.
cp -r models/* ../models
Prepare custom dataset. Place all your frames in one directory, following the naming convention:
<video_name>_<frame_number>.jpg
We have used publicly available UCF-101 dataset, which can be found here: https://www.crcv.ucf.edu/data/UCF101.php
Colonoscopy data cannot be publicly shared, but we allow datasets to be available upon request. For further informaction please contact Tomasz Urbańczyk at [email protected]
Train FrameSeedGenerator and VideoDiscriminator with pre-trained ProGAN.
python train.py -d <path_to_dataset>