Author here: I work at Plumerai, a startup that’s building chips for efficient inference of binarized neural network (BNN). We previously open-sourced Larq [1], the library we use to build and train BNNs, and Larq Zoo [2], our repository of pre-trained BNNs from the literature. Now we’re opening up the deployment side as well with Larq Compute Engine (LCE).

You can grab a BNN from Larq Zoo (or build and train your own using Larq), use the MLIR converter in LCE to convert it into a TensorFlow Lite-compatible Flatbuffer file, and then use the LCE runtime to run inference on mobile and edge devices like Android phones [3] and the Raspberry Pi [4] (support for Armv8-M microcontrollers coming soon).

Internally at Plumerai, LCE has enabled our ML researchers to benchmark different BNN architectures and downstream application implementations on real hardware, so we’re excited to see what other developers and researchers will use it for!

Happy to answer any questions :)

[1] docs: https://docs.larq.dev, github: https://github.com/larq/larq

[2] docs: https://docs.larq.dev/zoo, github: https://github.com/larq/zoo

[3] docs: https://github.com/larq/compute-engine/blob/master/docs/quic...

[4] docs: https://github.com/larq/compute-engine/blob/master/docs/buil...