Azure ML Pipeline for PyTorch, ONNX, and C Compilation

To see a summary discussion on other ML @ Edge tooling, see edge-ml-options.md

This repository contains a Azure ML pipeline implementation for:

1. Training a PyTorch model
2. Exporting to ONNX format
3. Converting ONNX to C code using onnx2c
4. Compiling and testing the C code
5. Building a minimal binary for deployment

Project Structure

.
├── README.md
├── diagrams
├── environments
│   ├── gcc
│   │   └── Dockerfile
│   ├── onnx2c
│   │   └── Dockerfile
│   └── pytorch
│       ├── Dockerfile
│       └── requirements.txt
├── requirements.txt
├── setup_pipeline.py
└── src
    ├── compile_test
    │   ├── run.py
    │   └── templates
    │       ├── model_impl.c
    │       ├── test_model.c
    │       └── time_series_model.h
    ├── minimal_binary
    │   ├── README.md
    │   ├── binary-size-guide.md
    │   ├── run.py
    │   └── templates
    │       ├── README.md
    │       ├── compile_minimal.sh
    │       ├── minimal_example.c
    │       ├── model_impl.c
    │       ├── nn_wrapper.h
    │       └── time_series_model.h
    ├── onnx2c
    │   └── run.py
    └── pytorch_train
        └── run.py

Prerequisites

• Azure ML workspace
• Azure CLI installed
• Python environment with the following packages:
  • azure-ai-ml
  • azure-identity

Setup Instructions

1. Clone this repository
2. Update setup_pipeline.py with your Azure ML workspace details (current setup uses a .env file)
3. Run the setup script:

# Set up the pipeline without running it
python setup_pipeline.py

# Or set up and immediately run the pipeline
python setup_pipeline.py --run

The --run flag will submit the pipeline job to Azure ML immediately. Without this flag, the setup script will only prepare the pipeline but not execute it.

Pipeline Components

1. PyTorch Training

• Trains a simple time series neural network using PyTorch
• Exports the model to ONNX format
• Saves test data for later validation

2. ONNX to C Conversion

• Uses onnx2c to convert the ONNX model to C code
• Creates additional C files needed for compilation and testing

3. C Compilation and Testing

• Compiles the C code generated from the model
• Runs tests using the test data saved during training
• Saves test results for analysis

4. Minimal Binary Build

• Creates a minimal binary suitable for embedded deployment
• Optimizes for size using compiler flags
• Provides memory usage statistics

Environment Details

PyTorch Environment

Contains PyTorch and all dependencies needed for training and ONNX export.

ONNX2C Environment

Contains the onnx2c tool built from source, which converts ONNX models to C code.

GCC Environment

Contains GCC and tools needed to compile and test the C code.

Pipeline Outputs

The pipeline produces several artifacts:

1. Training Output

   • ONNX model
   • Test data (CSV files)
   • Training metrics and visualizations

2. C Code Output

   • Generated C code from the ONNX model
   • Supporting C files for compilation

3. Test Results

   • Test output showing prediction accuracy
   • Compiled test binary

4. Minimal Binary

   • Optimized binary for deployment
   • Size and memory usage statistics

Customization

To adapt this pipeline for your own models:

1. Modify src/pytorch_train/run.py to train your specific model
2. Adjust the C wrapper code in src/onnx2c/run.py if your model has a different interface
3. Update compilation settings in src/minimal_binary/run.py for your target platform

Troubleshooting

• Check the logs for each pipeline step in the Azure ML Studio UI
• Ensure all environments are correctly built before running the pipeline
• Verify that the ONNX model is compatible with onnx2c (not all ONNX operations are supported)

License

MIT License