ClannXy/WFaceNet

PyTorch implementation for WFaceNet

WFaceNet

Introduction

• Train the model on CASIA-WebFace dataset, and evaluate on LFW dataset.

Requirements

• Python 3.6.10
• pytorch 0.4.1
• CUDA 9.2
• OpenCV-python
• scipy

Usage

Part 1: Preprocessing

• Download the aligned images at CASIA-WebFace@BaiduDrive(code:f9xx) and LFW@BaiduDrive(code:82gh).

Part 2: Train

1. Change the CASIA_DATA_DIR and LFW_DATA_DAR in config.py to your data path.

2. Train the WFaceNet model.

   Note: The default settings set the batch size of 128, use 1 gpu and train the model on 70 epochs. You can change the settings in config.py.

   python3 train.py
   

Part 3: Test

1. Test the model on LFW. train.py test the model on LFW after each train epoch complete automatically.

Results

• You can just run the lfw_eval.py to get the result, the accuracy on LFW like this:

    1    99.33
    2    99.33
    3    99.67
    4    98.83
    5    98.83
    6    99.67
    7    98.83
    8    99.50
    9    99.83
   10    99.67
   
   AVE    99.35

    python3 lfw_eval.py --resume --feature_save_dir
    ```
    * `--resume:` path of saved model
    * `--feature_save_dir:` path to save the extracted features (must be .mat file)

Reference resources

• arcface-pytorch
• MobileFaceNet_Pytorch