This repo contains pre-build Tensrflow and friends python wheels for various configurations

This is not the final wheel and requires a bit of manual setup. This wheel depends on Intel's MKL-DNN which needs to be installed either globally in your system (The standard locations for dynamic libraries are ~/lib, /usr/local/lib, and /usr/lib) or inside the site-packages/tensorflow.

This wheel was built with Conda's Intel MKL optimized numpy and C++ version of protobuf. I have tested this wheel only in the following configuration

1. Create conda env conda create -n mkl_tflow -y python=3.6 numpy protobuf==3.4.1
2. Activate it conda activate mkl_tflow
3. Install Tensorflow from wheel pip install https://github.com/anton-matosov/tensorflow-wheels/raw/master/Tensorflow-Wheels/MacOS/MKL/tensorflow-1.5.0-cp36-cp36m-macosx_10_7_x86_64.whl
4. Download https://github.com/intel/mkl-dnn/releases/download/v0.12/mklml_mac_2018.0.1.20171227.tgz and extract it in temporary location
5. Install libmklml.dylib and libiomp5.dylib into the library search path (e.g. ~/lib) or into your tensorflow's site-package folder ~/miniconda3/envs/mkl_tflow/lib/python3.6/site-packages/tensroflow (you can get site-package path by running python -c "from distutils.sysconfig import get_python_lib; print(get_python_lib())").
   You can use install_mkl.py script for this. Make sure to run it within the target conda env. This command will download and run the script: curl https://raw.githubusercontent.com/anton-matosov/tensorflow-wheels/master/Tensorflow-Wheels/MacOS/MKL/install_mkl.py | python -
6. Check that it all works by running python -c "import tensorflow as tf"
7. For maximum performance with MKL build you have to tweek environment variables for your specific CPU. Please refer to Optimizing for CPU - TensorFlow with Intel® MKL DNN page for details.
   Make sure to read main Optimizing for CPUas well. Fine tuning for your CPU is simple and always gives performance boost. In my case it was additional 1.5x on top of 4x provided by MKL, which is 6x total compared to naive CPU setup.

That's it, Enjoy AI!