Neural networks

[odow]

Prior art

ONNX

https://onnx.ai/about.html

OMLT

Supports ONNX

https://github.com/cog-imperial/OMLT/tree/main/src/omlt/neuralnet
https://github.com/cog-imperial/OMLT/blob/main/src/omlt/io/onnx_parser.py

Gurobi

Does not support ONNX

Use explicit bindings to PyTorch and tensorflow:

• keras https://github.com/Gurobi/gurobi-machinelearning/blob/main/src/gurobi_ml/keras/keras.py
• PyTorch https://github.com/Gurobi/gurobi-machinelearning/blob/main/src/gurobi_ml/torch/sequential.py

Supports only sequential NNs with ReLU and linear layers

Julia options

ONNX.jl

Seems partially supported. Not well maintained. The issue is the diversity of possible input models and their mapping to Julia.

Flux

Popular, if heavy, choice. Lots of options.

Lux

Well maintained. Simple. No ONNX support?

[odow]

#9 adds:

• Basic Lux support
• Linear layers
• ReLU

[odow]

We're missing ONNX support, but I'm going to close this until we pin down exactly how people will be using this library in Julia. It might be easier to use something like PyCall.