Skip to content

Latest commit

 

History

History
88 lines (63 loc) · 4.8 KB

README.md

File metadata and controls

88 lines (63 loc) · 4.8 KB

Tools Directory

Below is an overview of the scripts available in this directory and a brief description of their purpose:

Script Description
ns_autodeploy.py Deploys TFLite models to an EVB including binary creation, profiling, power measurement, and library generation.
ns_tflite_analyze.py Analyzes TFLite models to estimate MAC counts, memory reads/writes, and layer statistics; outputs reports in CSV/Excel.
generic_data.py eRPC server/client for handling generic data transfer (audio, sensor data, etc.) between EVB and PC.
ns_ad_batch.py Batch deployment of multiple models using YAML configuration files.
ns_test.py Automated testing framework for NeuralSPOT using configuration files and command-line arguments.
measurer.py Measures power consumption (using a Joulescope or similar) and sends data via ZMQ for visualization or logging.
opus_receive.py Receives and decodes Opus-encoded audio data from the EVB and writes decoded PCM data to a WAV file.
plotter.py Dash-based web application that displays live performance and power metrics graphs using Plotly.

Overview

This directory contains a collection of utility scripts that support various tasks for the NeuralSPOT project. These tools facilitate data communication between an EVB (Evaluation Board) and a PC, model analysis and deployment, automated testing, power measurement, and live data visualization. Below is an overview of each script and instructions on how to use them.

Contents

  • ns_autodeploy.py
    The primary tool for deploying TFLite models to an EVB board. This script supports:

    • Creating validation binaries.
    • Profiling model performance.
    • Power measurement (via Joulescope or onboard performance capture).
    • Generating minimal static libraries and AmbiqSuite examples. See the autodeploy Theory of Operations for more details.

  • ns_tflite_analyze.py
    Analyzes TensorFlow Lite (TFLite) model files find optimization targets, estimate MAC counts, memory reads/writes, and other layer statistics. It generates human-readable output in table format, as well as CSV and Excel reports.

  • ns_test.py
    Runs automated tests for the NeuralSPOT project. It uses a configuration file (INI format) along with command-line parameters parsed via pydantic_argparse to generate and run tests.

  • generic_data.py
    Implements an eRPC server and client for handling generic data operations between the EVB and PC. This tool supports:

    • Audio capture (32-bit stereo or 16-bit mono PCM) and writing to WAV files.
    • MPU6050 sensor data capture and logging to CSV.
    • Remote procedure calls for data fetching, computation (e.g., MFCC calculation), and printing.

  • ns_ad_batch.py
    Batch deploys multiple neural network models defined in a YAML configuration file. For each model, it builds the appropriate ns_autodeploy command and executes it with support for retries and logging. See the ad_batch Usage Guide for more details

  • measurer.py
    Measures power consumption (using a Joulescope or similar power measurement tool) during model inference. The script packs the measurement data using MessagePack and sends it over a ZMQ PUB socket (port 5556) for further processing or visualization.

  • opus_receive.py
    Receives Opus-encoded audio data sent over a serial connection via eRPC, decodes the Opus frames into PCM, and writes the output to a WAV file.

  • plotter.py
    A Dash-based web application that subscribes to power and performance data (via ZMQ) and displays live graphs (e.g., latency and energy consumption). It uses Plotly for real-time data visualization.

Prerequisites

ns_autodeploy can be installed using pip:

cd ..
pip install .

To run the other tools, ensure that you have Python 3.6 or later installed along with the following packages:

  • Core Packages:
    numpy, pyyaml, pydantic, pydantic_argparse

  • Audio & Signal Processing:
    soundfile, pyogg

  • Communication & RPC:
    erpc

  • Data Packing & Messaging:
    msgpack, msgpack_numpy

  • Visualization:
    dash, plotly

You can install the required packages using pip. For example:

pip install numpy pyyaml pydantic pydantic_argparse soundfile pyogg erpc msgpack msgpack_numpy dash plotly