IRDS Gesture Recognition Project

PyTorch-based gesture recognition system with multiple model architectures for skeleton data analysis.

🚀 Quick Start

cd model/train
export HIP_VISIBLE_DEVICES=1
python train_pytorch_only.py --config ../../config/experiment_medium.yaml

📚 Documentation

All documentation is located in the readme/ folder:

Getting Started

• QUICK_START.md - Fast reference guide for running experiments

Model & Training

• README_EXPERIMENTS.md - Comprehensive guide to multi-model experiments
• MULTI_MODEL_SETUP.md - Technical details of model architecture system

Cleanup & Maintenance

• CLEANUP_SUMMARY.md - Training directory cleanup summary
• MODEL_CLEANUP_SUMMARY.md - Model directory cleanup summary

Inference & Testing

• VIDEO_PIPELINE_GUIDE.md - Video-to-gesture inference guide

Benchmarking

• BENCHMARKING_GUIDE.md - Complete GPU benchmarking guide (AMD vs NVIDIA)
• BENCHMARK_SUMMARY.md - Quick reference for benchmarking

HPC Deployment

• HPC_FILE_LIST.md - Files needed for HPC deployment

📁 Project Structure

irds/
├── model/
│   ├── clip_gesture_model_pytorch.py    # Main model implementation
│   ├── model_architectures.py           # 6 model size variants (150K-51M params)
│   ├── train/                           # Training scripts
│   │   ├── train_pytorch_only.py        # Single GPU training
│   │   ├── train_pytorch_multi_gpu.py   # Multi-GPU training
│   │   ├── run_experiments.sh           # Run all experiments
│   │   └── compare_experiments.py       # Compare results
│   └── test/                            # Testing utilities
│       ├── video_to_gesture.py          # Video inference
│       └── test_pytorch_clean.py        # Environment tests
├── config/                              # Experiment configurations
│   ├── experiment_tiny.yaml             # ~150K params
│   ├── experiment_small.yaml            # ~1.1M params
│   ├── experiment_medium.yaml           # ~2.5M params (recommended)
│   └── experiment_large.yaml            # ~14M params
├── data/                                # Dataset directory
├── readme/                              # All documentation
└── utils.py, tensor_utils.py            # Utility functions

🎯 Available Models

Model	Parameters	Best For
tiny	150K	Quick experiments
small	1.1M	Simple patterns
medium	2.5M	Recommended
large	14M	Complex patterns
xlarge	24M	Very large datasets
xxlarge	51M	Extreme capacity

⚡ Common Commands

Run Single Experiment

cd model/train
python train_pytorch_only.py --config ../../config/experiment_medium.yaml

Run All Experiments

cd model/train
./run_experiments.sh

Compare Results

cd model/train
python compare_experiments.py

Multi-GPU Training

cd model/train
python train_pytorch_multi_gpu.py --config ../../config/experiment_medium.yaml --gpu 1

Benchmark Performance

# Run complete benchmark (training + inference)
bash scripts/run_full_benchmark.sh

# Compare AMD vs NVIDIA
python scripts/compare_gpu_benchmarks.py outputs/benchmark_amd_*.json outputs/benchmark_nvidia_*.json

🖥️ HPC Deployment

# Copy to HPC
rsync -avz --exclude='__pycache__' --exclude='*.pyc' --exclude='.git' \
    /home/nishant/project/irds/ user@hpc:/path/to/irds/

# On HPC
export HIP_VISIBLE_DEVICES=1
cd /path/to/irds/model/train
./run_experiments.sh

📊 Output Files

Trained models are saved in model/train/outputs/:

• model_<size>.pth - Model weights
• scaler_<size>.json - Feature scaler
• info_<size>.json - Training metrics
• curves_<size>.png - Training plots
• confusion_<size>.png - Confusion matrix

🔧 Requirements

• Python 3.8+
• PyTorch 2.0+ (with ROCm support for AMD GPUs)
• pandas, scikit-learn, pyyaml
• For video inference: opencv-python, mediapipe

📖 More Information

See the readme/ folder for detailed documentation on all aspects of the project.

---

Last Updated: October 27, 2025