🏋️ PoseAI Trainer

A web application that analyzes squat form using a custom-trained deep learning model and biomechanical pose metrics.

---

How It Works

1. Train – Upload a Google Drive folder containing labeled squat videos (good/ and bad/ subfolders). The app trains a 3D CNN video classifier on them.
2. Analyze – Upload any squat video and get an instant classification (Good / Needs Improvement) with two personalized coaching tips.

The model only trains once. After training, the app switches automatically to analysis mode.

---

Project Structure

PoseAITraining_web/
├── app.py          # Streamlit UI
├── main.py         # Model architecture, training pipeline
├── backend.py      # Video analysis, pose metrics (MediaPipe), feedback
├── outputs/        # Saved model files (auto-created after training)
├── .streamlit/     # Streamlit config
├── .env            # Environment variables (not committed)
└── requirements.txt

---

Getting Started

1. Install dependencies

pip install -r requirements.txt

2. Run the app

streamlit run app.py

3. Train the model

Paste a Google Drive folder link that follows this structure:

root/
  good/
    video1.mp4
    ...
  bad/
    video2.mp4
    ...

Click Train model and wait. Training runs once — the model is saved to outputs/pose_model.pt.

4. Analyze a video

Upload a squat video (MP4, MOV, AVI, MKV, WEBM) and click Analyze video.

---

Feedback System

Coaching tips are generated from two sources:

Source	When used
Biomechanical rules (MediaPipe pose metrics)	Always
OpenAI GPT-4o-mini	Only if OPENAI_API_KEY is set in .env

Metrics measured: squat depth, torso lean angle, knee tracking, left/right symmetry, movement stability.

---

Environment Variables (.env)

OPENAI_API_KEY=sk-...   # Optional – enables AI-generated feedback
EPOCHS=4                # Training epochs (default: 4)
BATCH_SIZE=2            # Training batch size (default: 2)
LEARNING_RATE=0.0001    # Learning rate (default: 1e-4)

---

Requirements

• Python 3.9+
• PyTorch
• OpenCV
• MediaPipe
• Streamlit
• gdown

---

Notes

• MediaPipe may fail when the .venv path contains non-ASCII characters (e.g. Hebrew folder names). In that case, pose metrics are skipped and classification still works normally.
• The model file (outputs/) is excluded from version control.