Docker-based build environment that creates Tensorflow packages that don't use optimised CPU instructions such as AVX, AVX2, SSE4.1, SSE4.2 and FMA. This is the version of Tensorflow that is used in Photonix to maintain compatibility with as many people's machines as possible. Performance testing detailed here showed a 13% longer execution time than the official, optimised builds.
Running the commands below should build Python Wheel packages for each supported architecture in the wheels directory. Beware that building will probably take many hours. We usually create the largest possible VPS on DIgitalOcean with 32 CPU cores and it can build each architecture in about an hour. Check out the releases page to make use of our builds.
./build.sh amd64
./build.sh arm32v7
./build.sh arm64v8
- Provide options for passing in custom PyPI server credentials and having wheels automatically uploaded using pypi-uploader