Forked from taotaonice fantastic job on FaceShifter. Made the following modifications:
- Moved from Visdom to Tensorboard to monitor the various loss contributors
- Moved to Pytorch native AMP implementation (Pytorch>=1.6): less memory consumption, increased batch size,
- Various corrections: generator hinge loss, dataset logic,...
Added DiffAugment algorithm to increase the discriminator perceived diversity,- Tested gradient accumulation, replacing Batch norm with GroupNorm: unsuccessful due to poor attributes transfer,
- Reduced the adversarial loss and the Generator learning rate to achieve better source id transfer,
- Changed the face cropping algorithm to get a better chin coverage.
Rework the dataset generation to keep FFHQ native images and to crop the others images using FFHQ algorithm (using landmarks) to preserve the whole chin and achieve better alignment.
conda create -n FaceShifter -c pytorch -c conda-forge 'pytorch>=1.6' torchvision tensorboard opencv cudnn
conda activate FaceShifter
git clone https://github.com/ocastan/FaceShifter
cd FaceShifter
Go to https://github.com/TreB1eN/InsightFace_Pytorch and download model_ir_se50.pth to face_modules directory.
Get face sources (you can look here https://github.com/mindslab-ai/faceshifter#preparing-data for datasets)
cd face_modules
python preprocess_images.py unarchived_source_directory cropped_faces_destination_directory
Modify train_face_sources in train_AEI.py accordingly.
python train_AEI.py
You may reduce or increase the batch_size in train_AEI.py according to your graphic card memory.
Monitor losses and generated images, running from another terminal:
tensorboard --logdir runs/