face_recognition

For training was use the follow dataset CelebA

• Face Alignment checkpoint. (Used dllib)
• Base model EfficientNet_V2_S
  • Checkpount trained on ImageNet-1K

Project structure:

• ./model - you can take all checkpoints from here

• jupyter/train_baseline.ipynb - baseline face recognition solution, no tricks(no scheduler, no augmentation, not all dataset, only 10 epoch) - test accuracy 78%

• face_alignment.py - face alignment

• jupyter/test_identification_rate_metric.ipynb - tests for Identificaton rate metric (TPR@FPR).

  • jupyter/train_triplet_margin_loss.ipynb - checked IR metric on triplet loss function

• jupyter/train_arcface.ipynb - trained arcface, was use pytorch-metric-learning

• jupyter/train_triplet_loss.ipynb - trained triplet loss, was use open-metric-learning

• app/ - here is simple demo of face recognition system

  • If you want run app locally, you can just change workdir to cd ./app and use docker
    • docker compose build
    • docker compose up

• dataset.py - some helpers for CelebA dataset