How yelboudouri/emonext works
Project type and purpose
• This is a deep‐learning research codebase for training a facial emotion recognition model (EmoNeXt) built on top of the ConvNeXt architecture. It’s neither a web service nor a microservices system—it’s a machine‐learning training pipeline.
File structure
• Top‐level files and directories:
• README.md → project description and quick start
• requirements.txt → Python dependencies (PyTorch, CUDA, etc.)
• models.py → model definitions (ConvNeXt backbone + STN + SE + self‐attention reg)
• train.py → main training loop, argument parsing, logging, checkpointing
• scheduler.py → learning‐rate scheduler or custom training scheduler logic
• scripts/download_dataset.py → data‐download utility for FER2013
README for additional insights
• Key components:
• Data: FER2013, downloaded via a helper script
• Model: Adapted ConvNeXt with Spatial Transformer Network and Squeeze‐and‐Excitation blocks
• Training features: mixed‐precision (AMP), configurable batch size, epochs, learning rate, model size
• Dependencies: CUDA, PyTorch 1.13+, any torchvision or auxiliary libs in requirements.txt