Tflite Nightly 2022-01-04 experimental

tflite-nightly-2022-01-04

Add readme

Tensorflow Lite Dec-19 nightly eb80c2b71 Build for Mac and Linux

nightly-eb80c2b71

Add script to package up tflite

Tensorflow Lite 2.6.2 Build for Mac and Linux

Supports Apple M1 and Linux AMD65 currently. Instructions here.

Tensorflow Lite 2.6.2 Build

Supports only Apple M1 currently. Instructions here.

tflite-nightly-2022-02-28

Optimized for Intel NUC11, no MacOS

Tensorflow Nightly Python Wheel macOS x86 (M1 compatible) Feb-2022

Feb-2022 tf-nightly (2.9) release, x86build compatible with M1.
version info v1.12.1-71751-g965c39fdf30 2.9.0 (from python -c "import tensorflow as tf; print(tf.__git_version__, tf.__version__)")

Tensorflow Nightly Python Wheel macOS x86 (M1 compatible) Jan-2022

Python Vendorized Files

We vendorize Tensorflow nightly builds here because there are currently no official builds that run reliably on Apple M1 chips.

macOS wheel

Note that we are running x86_64 binaries even on the M1, but we have to make sure they are compiled without incompatible instructions (see below).

We compile the M1 version ourselves:

1. check out the nightly branch from tensorflow
2. create a conda environment for the build process; make sure the correct version of Python as well as Numpy is installed; for example for a build on 2021-01-15 with Python 3.9, numpy version 1.21.5 was required; anything newer was incompatible with numba 🤷
3. activate the conda environment and initialize the bazel build environment with python configure.py (default answers everywhere are fine)
4. remove the line build:release_cpu_macos --config=avx_linux from the .bazelrc file to disable avx instructions that are not compatible with M1
5. modify the version string in tensorflow/tools/pip_package/setup.py to indicate a development version and the build date; for example change to _VERSION = '2.9.0.dev20220115'
6. build with bazel build //tensorflow/tools/pip_package:build_pip_package this can take up to several hours
7. create the wheel with ./bazel-bin/tensorflow/tools/pip_package/build_pip_package --nightly_flag /tmp/tensorflow_pkg
8. use the wheel in /tmp/tensorflow_pkg

Linux wheel

The Linux wheel can be downloaded from the pypi website or downloaded with pip; create a virtualenv with the desired python version and download with for example: python -m pip download tf_nightly==2.9.0.dev20220115. The wheel will reside in the current directory.

TF Nightly XNNPACK and GPU test 4fe7e603a0

TF Nightly XNNPACK and GPU test 4fe7e60