jjangsangy / MacOS-TensorflowBuilds

Optimized tensorflow wheels binaries build for macos

MacOS TensorflowBuilds

Optimized Tensorflow Binaries for MacOS

Serving the underserved

1. (Intel MKL-DNN 2018)

Version	Release	Optimization	OS
v1.3.1	2.7, 3.6	MKL MSSE4.2 MAVX2 MAVX	High Sierra
v1.4.0rc0	2.7, 3.6	XLA MKL MSSE4.2 MAVX2 MAVX	High Sierra
v1.4.0	2.7, 3.6	XLA MKL MSSE4.2 MAVX2 MAVX	High Sierra
v1.4.1	2.7, 3.6	XLA MKL MSSE4.2 MAVX2 MAVX	High Sierra
v1.7.0rc1	2.7, 3.6	XLA MKL MSSE4.2 MAVX2 MAVX	High Sierra

Note v1.7.0rc1 is slow

The suggested version to run currently running on MKL is 1.4.1

Intel MKL-DNN

Intel MKL-DNN includes functionality similar to Intel(R) Math Kernel Library (Intel(R) MKL) 2017, but is not API compatible. We are investigating how to unify the APIs in future Intel MKL releases.

This release contains a range of performance critical functions used in modern image recognition topologies including Cifar*, AlexNet*, VGG*, GoogleNet* and ResNet* optimized for wide range of Intel processors.

Python 3 Install

    $ TF_URL='https://github.com/jjangsangy/MacOS-TensorflowBuilds/releases/download/1.4.1/tensorflow-1.4.1-cp36-cp36m-macosx_10_13_x86_64.whl'
    $ pip3 installl --upgrade --force-upgrade "$TF_URL"

Python 2 Install

    $ TF_URL='https://github.com/jjangsangy/MacOS-TensorflowBuilds/releases/download/1.4.1/tensorflow-1.4.1-cp27-cp27m-macosx_10_13_x86_64.whl'
    $ pip2 installl --upgrade --force-upgrade "$TF_URL"

Optimizations

• MKL: Intel Math Kernel Library
• XLA: Accelerated Linear Algebra
• MSSE4 etc.: SIMD Vectorization Extensions

Build It Yourself

    $ ./install.sh -h

        DESCRIPTION:
        Installer script

        USAGE:
        ./install.sh --py ${py_version} --tf ${tf_version} --mkl ${mkl_dir}

        OPTIONS:
        -h| --help: Print this help message andn exit
        -p| --py version: [default 3.6]
            Spcify version of python
        -t| --tf version: [default 1.4.0rc0]
            Specify version of tensorflow
        -m| --mkl path: [default /opt/intel]
            Specify mkl library path
        -i| --inplace [default off]
            Install tensorflow inplace right after compilation