Blowtorch

Blowtorch is a Python package that allows you to train machine learning models and run inference in pure Rust. This is done through a specifying the model once in a JSON file. Blowtorch then exports your specification into Rust and Python models. You can train the Python model as you prefer, and Blowtorch can be run again to share the weights to Rust.

An example application built with a predecessor of Blowtorch is ZipNet, which is a neural-network based compression algorithm run entirely in the browser. We built Blowtorch as we could not find any easily extensible machine learning frameworks that could be compiled to WebAssembly.

Advantages over similar packages

• Inference is in pure Rust, meaning your model can run anywhere that Rust runs. You can for example compile it to WebAssembly.
• New layers can be implemented very easily, as one just has to write a forward pass in Rust
• Training is completely in Python, meaning you can use whatever training procedures you like
• Complex networks can be built by splitting the architecture into simpler modules, which are combined together by some glue code

Our documentation can be found at https://blowtorch.readthedocs.io/en/latest/.

Features

• Export and import trained weights
• Implementations for the following layers:
  • Conv
  • ConvT
  • ReLU
  • GDN
  • iGDN
  • Flatten
  • Linear
• Easy-to-use example
• Possibilities to extend the framework
• Documentation for Python & Rust