TensorFlow Binarize Ops

This repository contains the binarized Dense/Conv2D ops develoepd by @jwfromm for TensorFlow and Keras.

Building

There are two ways to build these custom ops:

• If you are developing on TensorFlow itself, you can mount this repository within the TensorFlow source tree and use bazel to build the ops. Mount this repository at a location such as <tensorflow root>/tensorflow/core/user_ops/BinOps, then run ./tensorflow/core/user_ops/BinOps/bazel_build_command.sh to invoke bazel to build the necessary libraries. The libraries will be placed within the bazel-bin directory that bazel creates within the TensorFlow directory.

• If you have TensorFlow already installed through something like pip, use ./gcc_build_command.sh to build the libraries. The libraries will be placed within this directory.

Once the .so libraries are built, install them into /usr/local/lib and run sudo ldconfig. You can then place the python files binarize_layers.py and binarize_ops.py wherever is convenient for python to pick them up from your own code.

Testing

To test this code, use run_tests.sh, which will build the library, run a set of unit tests against the codebase. It is highly encouraged to use docker to provide the codebase with a uniform environment already containing such necessities like tensorflow and CUDA preinstalled.

To run tests on CPU-only kernels:

docker run -e NO_GPU=1 -ti -v $(pwd):/src -w /src tensorflow/tensorflow:latest-devel-py3 ./run_tests.sh

To run tests on GPU and CPU kernels, run:

nvidia-docker run -ti -v $(pwd):/src -w /src tensorflow/tensorflow:latest-devel-gpu-py3 ./run_tests.sh

Of course, if you don't want to use docker and instead just want to run the tests directly on your host system, you can do so simply by running ./run_tests.sh.