DistilBERT NLU for Cockpit Intents

This project implements a Natural Language Understanding (NLU) module for a cockpit environment using a fine-tuned DistilBERT model. It includes the complete workflow from training in Python to high-performance inference using C++ and ONNX Runtime.

Project Overview

The system classifies user commands into specific cockpit intents:

• lower_window
• raise_window
• set_temperature
• mute_media
• unmute_media

The repository contains both the Python scripts for model development and a C++ implementation for efficient deployment.

Directory Structure

• onnx_model/: Contains the exported ONNX model (model.onnx).
• onnx_runtime_c++/: C++ implementation using ONNX Runtime for inference.
  • my_tokenizer/: Main C++ application source code.
  • tokenizers-cpp/: Submodule/dependency for tokenization.
  • src/: Additional C++ utility/inference files.
• distilbert_cockpit_intents_2nd/: Saved artifacts for the fine-tuned Hugging Face model.
• Python Scripts:
  • fine_tune.py: Main script for training the model.
  • tokenize_data.py: Pre-processing script to tokenize datasets.
  • inference_cockpit_intents.py: Example inference using PyTorch/Transformers.
  • inference_onnx_intents.py: Example inference using ONNX Runtime in Python.

Python Workflow

Prerequisites

Ensure you have the necessary Python libraries installed:

pip install transformers datasets evaluate torch numpy onnxruntime

1. Training

To fine-tune the DistilBERT model on your dataset:

python3 fine_tune.py

This script loads the data from CSV files (training_data.csv, etc.), tokenizes it, trains the model, and saves the results to distilbert_cockpit_intents_2nd.

2. Inference

Using PyTorch:

To test the trained model using standard Hugging Face components:

python3 inference_cockpit_intents.py

Using ONNX (Python):

To test the exported ONNX model via Python:

python3 inference_onnx_intents.py

C++ ONNX Runtime Implementation

This section details how to build and run the C++ inference engine.

Prerequisites

• CMake (version 3.13 or higher)
• C++ Compiler (supporting C++17)
• Dependencies: The project expects onnxruntime-linux-x64-1.20.0 and tokenizers-cpp to be present in the onnx_runtime_c++ directory.

Build Instructions

1. Navigate to the project directory:

   cd onnx_runtime_c++/my_tokenizer

2. Create a build directory and compile:

   mkdir build
   cd build
   cmake ..
   make

Usage

Run the compiled pipeline executable:

./pipeline

The program will launch an interactive CLI where you can type commands (e.g., "lower the temperature to 20") and see the predicted intent and confidence scores.

Configuration

The build configuration is managed in onnx_runtime_c++/my_tokenizer/CMakeLists.txt. Ensure the paths to onnxruntime and tokenizers-cpp matches your directory layout if you move files around.