FaceCom Multi-task Learning Project

This project implements a multi-task learning solution for face analysis, combining gender classification and face recognition tasks using a shared CNN backbone.

├── src/
│   ├── dataset.py      # Dataset and data loading utilities
│   ├── model.py        # Neural network architecture
│   ├── trainer.py      # Training and evaluation logic
│   ├── main.py         # Main training pipeline
│   └── run.py          # Entry point script
├── data/               # Dataset directory
│   └── facecom/        # FaceCom dataset
├── checkpoints/        # Model checkpoints
└── README.md          # Project documentation

• Python 3.7+
• PyTorch
• torchvision
• albumentations
• scikit-learn
• pandas
• numpy
• matplotlib
• tqdm

Install dependencies:

pip install torch torchvision albumentations scikit-learn pandas numpy matplotlib tqdm

Place your FaceCom dataset in the data/facecom directory with the following structure:

data/facecom/
├── images/            # Image files
└── annotations.json   # Annotation file

1. Update the configuration in src/main.py if needed:

   • Adjust learning rate, batch size, epochs
   • Modify model architecture parameters
   • Change loss weights (alpha for gender, beta for identity)

2. Run the training:

python src/run.py

Optional arguments:

python src/run.py --data_dir path/to/data --batch_size 32 --epochs 30 --lr 0.001 --seed 42

• Backbone: ResNet-50 (pretrained)
• Two task-specific heads:
  • Gender Classification Head (Binary)
  • Face Recognition Head (Multi-class)

• Multi-task learning with shared backbone
• Comprehensive data augmentation for challenging visual conditions
• Early stopping and model checkpointing
• Detailed metric tracking and visualization
• Production-ready code structure

The model tracks the following metrics:

• Accuracy
• Precision
• Recall
• F1 Score

• Top-1 Accuracy
• Macro F1 Score

• Weighted combination (30% Gender, 70% Identity)

• Training metrics are plotted and saved as training_metrics.png
• Best model is saved in the checkpoints directory
• Predictions are saved in submission.csv