Start here if you are new to uTensor or TinyML. These tutorials contain step-by-step instructions to achieve specific tasks.
This repository contains a Jupyter training notebook and Mbed skeleton code to demonstrate the uTensor end-to-end flow.
The uTensor code-generator documentation is great for those who wish to build their own TinyML pipelines.
uTensor frontend parsers read files or objects and convert them to uTensor's graph representation, ugraph. Once the input graphs are converted into the ugraph format, existing optimizations and export methods can be applied to them.
A graph-transformer accepts an input ugraph, modifies it, and returns an output ugraph. It is used to realize functions such as node fusion/removal/replacement, quantization, and graph-lowering.
The backend of uTensor Code-Generator is responsible for converting final ugraphs to a deployable model representation. The uTensor C++ code-generation is an example of the backend. Outputting other formats, such as RTL, C, TensorFlow Lite model, are also possible.
The uTensor runtime documentation summarizes the design philosophy of the uTensor device-code. It also includes tutorials on how to extend the uTensor runtime.