Sensor Fusion for GPS Position Estimation

This repository contains code for training and evaluating a gated sensor fusion neural network for GPS position estimation. The project leverages PyTorch, Weights & Biases (wandb), and custom data loaders for handling sensor data.

Features

• Custom PyTorch datasets for loading and batching sensor fusion data
• Gated sensor fusion neural network model
• Training and validation loops with early stopping
• Inference pipeline for evaluating model performance
• Integration with Weights & Biases for experiment tracking

Project Structure

• dataloaders.py: Custom PyTorch Dataset classes for training and inference
• gated_fusion_net.py: Implementation of the Gated Sensor Fusion neural network
• train_loop.py: Training script with argument parsing, early stopping, and wandb logging
• Infererence.py: Inference script for evaluating the model on test data
• utils.py: Utility functions (e.g., plotting, reading folder names)
• lists/: Text files listing data folders for training, validation, and testing

Setup

1. Install dependencies:

   pip install -r requirements.txt

   Ensure you have torch, pandas, numpy, wandb, python-dotenv, and other required packages.

2. Set up Weights & Biases:

   • Create a .env file with your WANDB API key:

     WANDB_API_KEY=your_wandb_api_key_here
     

3. Prepare data:

   • Place your data folders and CSV files as referenced in the lists/ directory.

Usage

Training

Run the training loop with default or custom arguments:

python train_loop.py --train_list lists/train.txt --val_list lists/val.txt --seen_data_dir ./seen_subjects_test_set --output_dir ./no_drp_out_checkpoint

Inference

Run the inference script to evaluate the trained model:

python Infererence.py

Notes

• The model expects CSV files named synced_gps_gt_roin.csv in each data folder.
• Training and validation folder names should be listed in the corresponding text files in the lists/ directory.
• Results and logs are tracked using wandb.

License

This project is for educational and research purposes.