tensorflow.md

Tensorflow Lite in Faasm

Tensorflow Lite is well suited to running inference tasks in a resource-constrained serverless context.

It is written in C/C++ hence we can build Tensorflow Lite to WebAssembly using the standard Faasm toolchain.

Faasm currently only supports the C/C++ API, but building the Python API using Faasm's Python support should be possible and on the to-do list.

Compiling TensorFlow Lite to WebAssembly

To do this, make sure you've checked out this project and updated all the git submodules, then, from the Faasm CLI:

inv compile-tflite

The build output ends up at third-party/tensorflow/tensorflow/lite/tools/make/gen.

Building a C/C++ function with TF Lite

A function implementing image classification is included in the examples. This is based on the example in the TF Lite repo. The data and model for the example are stored in this repo.

To run the function, you need to run a local Faasm cluster (as described in the README), then:

# Upload files and state
inv tf-upload-data tf-state-upload

# Upload (takes a few seconds)
inv upload tf image

# Invoke
inv invoke tf image

Eigen Fork

To support WASM simd instructions I've hacked about with Eigen on a fork. It seems to work but isn't well tested.

TF Lite will be compiled against the version of Eigen downloaded as part of its 3rd party deps, so if you need to change it and rebuild you'll need to run:

cd third-party/tensorflow/tensorflow/lite/tools/make
rm -r downloads/eigen/
./download_dependencies.sh