LIM-RPS and DELL Evaluation Guide

LIM-RPS and DELL Evaluation Guide

This guide explains how to evaluate sensor data using the LIM-RPS and DELL algorithms with visualizations.

Overview

The `evaluate_with_limrps.py` script:
1. Loads sensor data from CSV files
2. Processes data through LIM-RPS (Lipschitz-constrained Implicit-Map for Recursive Proximal Synthesis)
3. Optionally processes through DELL (Dual-Equilibrium Latent Learning)
4. Generates comprehensive visualizations

Requirements

pip install torch numpy scipy matplotlib

Or use the full requirements:

pip install -r requirements.txt

Usage

Basic Evaluation with Visualizations

python3 evals/evaluate_with_limrps.py \
  --sensor-dir "/path/to/sensor/data" \
  --plots-dir evals/plots

With LIM-RPS Processing

python3 evals/evaluate_with_limrps.py \
  --sensor-dir "/path/to/sensor/data" \
  --use-limrps \
  --plots-dir evals/plots

With DELL Processing

python3 evals/evaluate_with_limrps.py \
  --sensor-dir "/path/to/sensor/data" \
  --use-limrps \
  --use-dell \
  --plots-dir evals/plots

Evaluate Specific Session

python3 evals/evaluate_with_limrps.py \
  --sensor-dir "/path/to/sensor/data" \
  --use-limrps \
  --session "Left Pocket" \
  --plots-dir evals/plots

Output

Visualizations

The script generates plots in the specified `--plots-dir` directory (default: `evals/plots/`):

1. Accelerometer X, Y, Z over time - Raw accelerometer data
2. Accelerometer Magnitude - L2 norm of acceleration vector
3. Motion Energy - Computed motion energy over time
4. Gyroscope Data - Angular velocity (if available)
5. LIM-RPS Convergence - Residual convergence plot (if LIM-RPS is used)
6. 3D Accelerometer Trajectory - 3D visualization of acceleration space

Each plot is saved as `{session_name}_analysis.png`.

Example Output Structure

evals/
├── plots/
│   ├── Left Pocket_2025-11-02_02-10-10-..._analysis.png
│   ├── Right Pocket_2025-11-02_02-10-10-..._analysis.png
│   └── Single_442_W_13th_St-2025-11-02_04-37-38_analysis.png
└── limrps_evaluation.md

What LIM-RPS Does

LIM-RPS performs multi-modal sensor fusion by:

1. Encoding: Converts raw sensor data (accelerometer, gyroscope, energy) into latent representations
2. Fusion: Solves a fixed-point equation to find the equilibrium state:

   z* = prox_G(z* - γ B_θ(z*, e))

where:
- `B_θ` is a 1-Lipschitz cross-modal operator
- `prox_G` is a proximal projection
- `γ` is the step size
3. Convergence: Iteratively refines the fused representation until convergence

Key Metrics

• Residuals: Measure of convergence at each iteration
• Fused Latents: Combined representation of all sensor modalities
• Convergence Rate: How quickly the algorithm converges

What DELL Does

DELL (Dual-Equilibrium Latent Learning) extends LIM-RPS with:

1. Fast Equilibrium: Frame-rate processing (60-100 Hz) for micro-adjustments
2. Slow Equilibrium: Beat-rate processing (2-4 Hz) for macro-planning
3. Coupling: Energy function `φ(x, y)` keeps fast and slow equilibria aligned

Understanding the Graphs

### Accelerometer X, Y, Z
- Shows raw acceleration in each axis
- Useful for identifying motion patterns
- Z-axis typically shows vertical motion

### Accelerometer Magnitude
- L2 norm: `√(x² + y² + z²)`
- Represents overall motion intensity
- Useful for detecting activity levels

### Motion Energy
- Computed from accelerometer features
- Normalized energy metric
- Higher values indicate more motion

### LIM-RPS Convergence
- Shows residual values at each iteration
- Should decrease (converge) over iterations
- Log scale shows convergence rate

### 3D Trajectory
- Visualizes acceleration in 3D space
- Shows motion patterns and orientations
- Useful for understanding movement geometry

Troubleshooting

### "torch not available"
Install PyTorch:

pip install torch

### "matplotlib not available"
Install matplotlib:

pip install matplotlib

### No convergence in LIM-RPS
- Check that sensor data is valid
- Try adjusting LIM-RPS parameters (step_size, max_iters)
- Ensure data is properly normalized

### Missing visualizations
- Check that matplotlib is installed
- Verify output directory is writable
- Check for errors in the console output

Advanced Usage

Custom LIM-RPS Configuration

Edit the script to customize LIM-RPS parameters:

config = LIMRPSConfig(
    max_iters=6,        # More iterations
    step_size=0.3,      # Smaller step size
    hidden_dim=128,     # Larger hidden dimension
    use_metric=True,    # Enable adaptive metric
    use_step_field=True # Enable adaptive step field
)

Processing Multiple Sessions

The script automatically processes all sessions found in the sensor directory. Use `--session` to filter specific sessions.

Integration with Main Evaluation

This script complements `evaluate_sensor_data.py`:
- `evaluate_sensor_data.py`: Basic statistics and metrics
- `evaluate_with_limrps.py`: Advanced fusion and visualizations

Use both for comprehensive analysis!