Human Gait Representation Learning

This project implements a Masked Autoencoder (MAE) for learning representations from IMU (Inertial Measurement Unit) data for human gait analysis.

Project Structure

• config.py: Configuration settings for the model
• dataset.py: Dataset loading and preprocessing
• mae.py: Implementation of the Masked Autoencoder model
• trainer.py: Training utilities
• training.py: Main training script
• utils.py: Helper functions
• visualize_output.py: Visualization tool for model input vs reconstruction
• evaluate_model.py: Evaluation metrics and performance analysis
• visualize_masking.py: Visualization of the masking process

Visualization Tools

The project includes three visualization tools to help understand the model:

1. Input vs Output Visualization

This tool allows you to visualize how well the model is reconstructing the original input data.

python visualize_output.py --checkpoint path/to/model/checkpoint.pth --data_path path/to/test/data --output_dir ./visualizations

Optional arguments:

• --num_samples 5: Number of samples to visualize
• --channels 0 1 2: Which channels to plot (defaults to first 3)
• --device cuda: Device to run on ('cuda' or 'cpu')

2. Model Evaluation

This tool provides quantitative evaluation of the model's performance.

python evaluate_model.py --checkpoint path/to/model/checkpoint.pth --data_path path/to/test/data --output_dir ./evaluation_results

Optional arguments:

• --batch_size 32: Batch size for evaluation
• --num_samples 100: Limit number of samples to evaluate (None = use all)
• --device cuda: Device to run on ('cuda' or 'cpu')

3. Masking Visualization

This tool helps visualize the masking and reconstruction process to understand what parts of the input the model is focusing on.

python visualize_masking.py --checkpoint path/to/model/checkpoint.pth --data_path path/to/test/data --output_dir ./masking_visualizations

Optional arguments:

• --num_samples 3: Number of samples to visualize
• --channels 0 1 2: Which channels to plot (defaults to first 3)
• --mask_ratio 0.75: Ratio of patches to mask
• --device cuda: Device to run on ('cuda' or 'cpu')

Training the Model

To train the model from scratch:

python training.py

Using a Pre-trained Model

If you have a pre-trained model, you can evaluate it or visualize its outputs using the provided tools.

Example:

# Evaluate the model
python evaluate_model.py --checkpoint /path/to/checkpoint.pth --data_path /path/to/test/data

# Visualize reconstructions
python visualize_output.py --checkpoint /path/to/checkpoint.pth --data_path /path/to/test/data

Project Details

This project uses a Masked Autoencoder architecture for self-supervised representation learning from time-series IMU data. The model is trained by:

1. Masking random parts of the input signal (default: 75% masked)
2. Encoding the visible portions with a Transformer encoder
3. Reconstructing the full signal with a Transformer decoder
4. Training on the reconstruction loss

The learned representations can be used for various downstream tasks related to human gait analysis.