nntrongnghia/mobile-arcface-torch

Face verification for mobile

Face Verification for Mobile

Train face verification model with ArcFace implemented with PyTorch. Models are intended to be used on mobile devices.

Install

This repo is tested with Python 3.7 and PyTorch 1.13. Other packages' versions can be found in requirements.txt.

Setup the dataset

This project use 2 datasets:

• MS1Mv3 (also called MS1M-RetinaFace) for training
• LFW (deep funneled) for validation

MS1Mv3 for training

• Download and extract the dataset here
• In the config file configs/ms1mv3_mbf.py, modify the config.root_dir value with the path to the extracted folder containing the MS1Mv3 dataset.

LFW for validation

• Download and extract the deep funneled version here. In the config file configs/ms1mv3_mbf.py, modify the img_dir value of config.lfwpair_kwargs with the path to the extracted folder containing LFW images.

• Download the pairs.txt here. In the config file configs/ms1mv3_mbf.py, modify the pairs_txt_path value of config.lfwpair_kwargs with the path to the pairs.txt file.

Train

Run python train.py. You can modify the config.batch_size in the config file to fit your GPU memory if needed.

Then run tensorboard --logdir ./lightning_logs to view the training progress in realtime.

The best checkpoint will be saved in the corressponding folder in ./lightning_logs/.

Metrics logged during training

• mean_target_cosine: Training metric - Mean value in each batch of cosine similarity between the image and the target class cluster center
• train_loss: ArcFace loss
• lfw_auroc: Validation metric - AUC ROC based on predictions of LFW pair images.
• pos_mean_score: Validation metric - Mean similarity score of all matching pair images
• neg_mean_score: Validation metric - Mean similarity score of all mismatch pair images