PyTorch and TensorFlow Co-Execution for Training a Speech Command Recognition System

This repo provides examples of co-executing MATLAB® with TensorFlow and PyTorch to train a speech command recognition system.

Signal processing engineers that use Python to design and train deep learning models are still likely to find MATLAB® useful for tasks such as dataset curation, signal pre-processing, data synthesis, data augmentation, and feature extraction. Open-source alternatives exist for those tasks and they could be OK to use when replicating a pre-existing model or training recipe. However, for original technical development work, most users find those tasks easier in MATLAB®.

Creator: MathWorks® Development

Requirements

• Python™
• MATLAB® R2021a or later
• Deep Learning Toolbox™
• Audio Toolbox™

To accelerate training, a GPU and the following toolbox is recommended:

• Parallel Computing Toolbox™

This repo includes two co-execution examples, with additional requirements.

CallMATLABFromPythonPytorch.mlx

• PyTorch (tested with version 1.9.0) and NumPy (tested with 1.21.1)
• MATLAB Engine API

CallPythonTensorFlowFromMATLAB.mlx

• TensorFlow (tested with version 2.0.0)
• Configured Python interpreter

Get Started

See SetupNotes.mlx for setup instructions for both examples included with this repo.

There are two high-level examples in this repo.

Call MATLAB from Python

CallMATLABFromPythonPytorch.mlx - In this example, Python™ is your main environment. You call into MATLAB® to perform dataset management and audio feature extraction.

Call Python from MATLAB

CallPythonTensorFlowFromMATLAB.mlx - In this example, MATLAB® is your main environment. The dataset management, audio feature extraction, training loop, and evaluation happen in MATLAB®. The deep learning network is defined and executed in Python™.

License

The license is available in the License file in this repository.