nengo-edge 24.3.6

Tools for working with NengoEdge

NengoEdge Tools

NengoEdge is a cloud-based platform for training and deploying high accuracy, low power audio AI models on edge devices. This package contains tools and examples to assist in taking a trained model exported from NengoEdge and deploying it in your own application.

To get started running NengoEdge models locally, set up a Python environment using the installation instructions below. Then download the live microphone demo notebook and open it with:

jupyter notebook /path/to/microphone-demo.ipynb

Installation

NengoEdge models use the TensorFlow machine learning library. If you already have TensorFlow installed, then all you need is to:

pip install nengo-edge

If you do not have TensorFlow installed, see the see the full installation instructions for more details.

Release history

24.3.6 (March 6, 2024)

Added

• Updated documentation with new content (additional tutorials, CLI documentation, FAQ section). (#9)

• Added support for input string processing and NLP model inference in SavedModelRunner. (#17)

• Added NetworkTokenizer class to perform remote calls to a device CLI that supports sentencepiece tokenization. (#17)

• Added stdio-based CLI runner for np_mfcc. (#19)

Changed

• SavedModelRunner tokenizer now uses SentencepieceTokenizer instead of FastSentencepieceTokenizer to ensure compatibility with the core sentencepiece library. (#17)

• Moved device_modules/network_tokenizer.py to network_runner.py. (#19)

Fixed

• Fixed model output decoding for ASR. SavedModelRunner now removes blank tokens and merges repeating tokens before detokenization. (#17)

Removed

• Removed support for streaming in SavedModelRunner. (#19)

23.9.27 (September 27, 2023)

Added

• Added warning when a downloaded nengo-edge model artifacts’ version does not match local environment nengo-edge version. (#6)

• Added nengo-edge package-dataset CLI command, which can be used to validate and package KWS and ASR datasets. (#10)

• SavedModelRunner.run now automatically decodes ASR model outputs via the exported sentencepiece tokenizer. (#15)

Changed

• SavedModelRunner now uses Keras’ SavedModel format, instead of the raw TensorFlow SavedModel format. (#8)

• SavedModelRunner can now take ragged object-arrays as input. (#8)

• TensorFlow is now an optional dependency installed with pip install nengo-edge[optional]. (#13)

23.7.30 (July 30, 2023)

Added

• Added CoralRunner for running models exported for the Coral board. (#4)

• Added DiscoRunner for running models exported for the Disco board. (#4)

• Added NordicRunner for running models exported for the Nordic board. (#4)

• Added on-device MFCC extraction code (device_modules.np_mfcc.LogMelFeatureExtractor). (#4)

• Added two new examples demonstrating how to run models exported for the Coral/Disco/Nordic devices. (#4)

Changed

• Renamed tflite_runner.Runner to TFLiteRunner. (#4)

• Renamed saved_model_runner.Runner to SavedModelRunner. (#4)

• TFLiteRunner.reset_state now takes a batch_size argument, which can be used to prepare the model to run with different batch sizes. (#5)

23.2.23 (February 23, 2023)

Fixed

• Fixed an issue causing pip to refuse to install nengo-edge. (#3)

23.1.31 (January 31, 2023)

Initial release