PyTorch implementation for BYHE.

1. Please install Requirements and download the VIPL-HR datasets.

   Requirements:
   python==3.7             
   torch==1.7.1+cu110       # the later version is also OK.
   torchvision==0.8.2+cu110
   numpy==1.19.5           
   pycwt==0.3.0a22
   opencv-python==4.5.4.60  
   scipy==1.7.2
   einops==0.3.2
   dlib==19.17.99
   tqdm==4.62.3
   scikit-learn==1.0.1
   matplotlib==3.3.3
   albumentations==1.2.0
   Pillow==8.3.2

   Note: It is recommended to use python 3.7, different versions will not be able to use the libraries we provide, details in ./libs.If so, contact us!

2. Preprocess VIPL-HR datasets. "vipl-hr-infos" can be found at ./data.

   python VIPL_preprocess.py 
       --data-path /path/to/vipl-hr/ 
       --infos-path ./data/vipl-hr-infos 
       --frame-path /path/to/save/frame_list 
       --mask-path /path/to/save/mask_list 
       --wave-path /path/to/save/wave_gt 
       --face-data-path /path/to/save/face_data 
       --face-img-path /path/to/save/face_img

   Note: If you don't need to pre-process the entire datasets, try modifying the version_type、person_name in VIPL_preprocess.py.

Default hyperparameter settings based on single GPU card "Tesla v100" with 16 GB memory.

python train.py 
    --frame-path /path/to/frame_list 
    --mask-path /path/to/mask_list 
    --wave-path /path/to/wave_gt 
    --GPU-id 0 
    --num-workers 4 
    --batch-size 6 
    --epochs 50 
    --log-enable True 
    --log-theme VIPL

Note: Default training fold 1, try modifying the test_set in train.py.

Pre-trained weights can be found at ./pretrained.

Test with your own data:

python test.py 
    --frame-path /path/to/frame_list 
    --mask-path /path/to/mask_list 
    --wave-path /path/to/wave_gt 
    --GPU-id 0 
    --num-workers 4 
    --log-enable True 
    --visual-enable True 
    --pretrained ./pretrained/VIPL_f1.pth

Note: Default testing fold 1, try modifying the test_set in test.py.

For quick testing:

python test.py 
    --frame-path ./data/vipl-frame/frame_list 
    --mask-path ./data/vipl-frame/mask_list 
    --wave-path ./data/vipl-frame/wave_gt 
    --GPU-id 0 
    --num-workers 4 
    --log-enable True 
    --visual-enable False 
    --pretrained ./pretrained/VIPL_f1.pth

This repository is released under the MIT license as found in the LICENSE file.