mlliarm / model2wasm

Convert your favourite machine learning models to WASM

Model to WASM

Convert a Python model to WASM object through the power of Python and Go.

Installation

To install, we need to install m2cgen. It is probably best to use a virtual environment such as venv to not pollute the global python setup, via something like python -m venv venv;venv/Scripts/activate;

pip install m2cgen

You will need to install the appropriate depedencies along with m2cgen depending on your model choice, e.g. pip install m2cgen scikit-learn

Also make sure you have Go and TinyGo installed. Optionally you may want to install just and wasmer

Usage

The rough workflow is:

1. Generate the model (in Go) via m2cgen
2. Generate the Go code
3. Generate the TinyGo compatible code for WASM build
4. Build the WASM file

It would look something like the below (has some missing arguments, check justfile for details:

$ m2cgen mymodel.pkl --language go > model2tinygo.go
$ go run model2tinygo.go > main.go
$ tinygo build -o output.wasm -target=wasi -wasm-abi=generic main.go
$ wasmer output.wasm -- 1 2 -2 -1

This has been abstracted away via just where you can run

just -l
just build-wasm mymodel.pkl mymodel.go mymodel.wasm
# in windows
just --shell powershell.exe --shell-arg -c build-wasm mymodel.pkl mymodel.go mymodel.wasm

Reproducible example:

python demo/generate_model_example.py --filename mymodel
just build-wasm mymodel.pkl mymodel.go mymodel.wasm
wasmer mymodel.wasm -- 1 2 -2 -1
# output - 199

Notes

Since I'm still getting around learning go there are going to be patterns which don't make sense to a seasoned programmer. One example is I don't understand go generate- does that help or hinder this pipeline?