This repository provides code for machine learning algorithms for edge devices developed at Microsoft Research India.
Machine learning models for edge devices need to have a small footprint in terms of storage, prediction latency, and energy. One instance of where such models are desirable is resource-scarce devices and sensors in the Internet of Things (IoT) setting. Making real-time predictions locally on IoT devices without connecting to the cloud requires models that fit in a few kilobytes.
Algorithms that shine in this setting in terms of both model size and compute, namely:
- Bonsai: Strong and shallow non-linear tree based classifier.
- ProtoNN: Prototype based k-nearest neighbors (kNN) classifier.
- EMI-RNN: Training routine to recover the critical signature from time series data for faster and accurate RNN predictions.
- Shallow RNN: A meta-architecture for training RNNs that can be applied to streaming data.
- FastRNN & FastGRNN - FastCells: Fast, Accurate, Stable and Tiny (Gated) RNN cells.
- DROCC: Deep Robust One-Class Classfiication for training robust anomaly detectors.
These algorithms can train models for classical supervised learning problems with memory requirements that are orders of magnitude lower than other modern ML algorithms. The trained models can be loaded onto edge devices such as IoT devices/sensors, and used to make fast and accurate predictions completely offline.
A tool that adapts models trained by above algorithms to be inferred by fixed point arithmetic.
- SeeDot: Floating-point to fixed-point quantization tool.
Applications demonstrating usecases of these algorithms:
- GesturePod: Gesture recognition pipeline for microcontrollers.
- MSC-RNN: Multi-scale cascaded RNN for analyzing Radar data.
- The
tf
directory contains theedgeml_tf
package which specifies these architectures in TensorFlow, andexamples/tf
contains sample training routines for these algorithms. - The
pytorch
directory contains theedgeml_pytorch
package which specifies these architectures in PyTorch, andexamples/pytorch
contains sample training routines for these algorithms. - The
cpp
directory has training and inference code for Bonsai and ProtoNN algorithms in C++. - The
applications
directory has code/demonstrations of applications of the EdgeML algorithms. - The
tools/SeeDot
directory has the quantization tool to generate fixed-point inference code.
Please see install/run instructions in the README pages within these directories.
For details, please see our project page, Microsoft Research page, the ICML '17 publications on Bonsai and ProtoNN algorithms, the NeurIPS '18 publications on EMI-RNN and FastGRNN, the PLDI '19 publication on SeeDot compiler, the UIST '19 publication on Gesturepod, the BuildSys '19 publication on MSC-RNN, the NeurIPS '19 publication on Shallow RNNs, and the ICML '20 publication on DROCC.
Also checkout the ELL project which can provide optimized binaries for some of the ONNX models trained by this library.
Code for algorithms, applications and tools contributed by:
- Don Dennis
- Yash Gaurkar
- Sridhar Gopinath
- Sachin Goyal
- Chirag Gupta
- Moksh Jain
- Ashish Kumar
- Aditya Kusupati
- Chris Lovett
- Shishir Patil
- Oindrila Saha
- Harsha Vardhan Simhadri
Contributors to this project. New contributors welcome.
Please email us your comments, criticism, and questions.
If you use software from this library in your work, please use the BibTex entry below for citation.
@software{edgeml03,
author = {{Dennis, Don Kurian and Gaurkar, Yash and Gopinath, Sridhar and Goyal, Sachin
and Gupta, Chirag and Jain, Moksh and Kumar, Ashish and Kusupati, Aditya and
Lovett, Chris and Patil, Shishir G and Saha, Oindrila and Simhadri, Harsha Vardhan}},
title = {{EdgeML: Machine Learning for resource-constrained edge devices}},
url = {https://github.com/Microsoft/EdgeML},
version = {0.3},
}
Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.