tensorflow intel platform 优化

intel平台优化

TensorFlow *是深度学习领域中主要使用的机器学习框架，要求高效利用计算资源。 为了充分利用英特尔架构和提高性能，TensorFlow *库已经使用英特尔MKL-DNN原语进行了优化，该原语是深度学习应用的流行性能库。

已进行优化的平台

有三种安装方式。

1. 使用pip

pip install -i https://pypi.anaconda.org/intel/simple tensorflow

2. anaconda 安装

video

3.  自己编译

前两种方式可能不支持最新的指令集。

首先安装 dnf Bazel

安装 Bazel

pushd /var/tmp

URL=https://github.com/bazelbuild/bazel/releases/latest
LASTURL=$(curl $URL -s -L -I -o /dev/null -w '%{url_effective}')
BZ_VERSION=${LASTURL##*/}
wget https://github.com/bazelbuild/bazel/releases/download/$BZ_VERSION/bazel-$BZ_VERSION-installer-linux-x86_64.sh

chmod +x bazel-*
./bazel-*
export PATH=/usr/local/bin:$PATH

popd

centos 7.4 can not install `dnf`from epel

WARNING: EPEL 7 DNF is very old and has issues to include security flaws. This appears to be the reason it was removed. That said here is the work around to get it working on Centos 7.

cat  >  /etc/yum.repos.d/dnf-stack-el7.repo << EOF
[dnf-stack-el7]
name=Copr repo for dnf-stack-el7 owned by @rpm-software-management
baseurl=https://copr-be.cloud.fedoraproject.org/results/@rpm-software-management/dnf-stack-el7/epel-7-\$basearch/
skip_if_unavailable=True
gpgcheck=
gpgkey=https://copr-be.cloud.fedoraproject.org/results/@rpm-software-management/dnf-stack-el7/pubkey.gpg
enabled=
enabled_metadata=
EOF

yum install dnf

centos 7会出现这个bug:

dnf copr plugin not present in dnf-plugins-core

因为EPEL 7 DNF 已经被移除了centos 7 install dn，还需要：

wget http://springdale.math.ias.edu/data/puias/unsupported/7/x86_64/dnf-plugins-core-0.1.5-3.sdl7.noarch.rpm

dnf install copr-cli

sudo dnf update

dnf copr enable vbatts/bazel

centos 可以直接安装bazel下：

wget https://copr.fedorainfracloud.org/coprs/vbatts/bazel/repo/epel-7/vbatts-bazel-epel-7.repo -P /etc/yum.repos.d/
yum install dnf-plugins-core-0.1.-.sdl7.noarch.rpm
yum install bazel 

install tf:

git clone https://github.com/tensorflow/tensorflow tensorflow
cd tensorflow

Compiling TensorFlow with Intel C Compiler

CC=icc bazel build --verbose_failures --config=mkl --copt=-msse4.2 --copt="-DEIGEN_USE_VML" -c opt //tensorflow/tools/pip_package:build_pip_package

bazel build --config=mkl -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mavx512f --copt=-mavx512dq --copt=-mavx512cd --copt=-mavx512bw --copt=-mavx512vl --copt="-DEIGEN_USE_VML" //tensorflow/tools/pip_package:build_pip_package

Build and Install TensorFlow* on Intel® Architecture

build tensorflow container：

more @ github

ref build-dev-container.sh  @github tensorflow docker

# source tf-docker.evn

# cat tf-docker.evn

# The script set the following environment variables for tf docker:
export TF_DOCKER_BUILD_TYPE=mkl
# export TF_DOCKER_BUILD_TYPE=CPU
# CPU or GPU image

export TF_DOCKER_BUILD_IS_DEVEL=YES
# Is this developer image

export TF_DOCKER_BUILD_DEVEL_BRANCH=r1.
# export TF_DOCKER_BUILD_DEVEL_BRANCH=master
#  (Required if TF_DOCKER_BUILD_IS_DEVEL is YES)
#  Specifies the branch to checkout for devel docker images

# export TF_DOCKER_BUILD_CENTRAL_PIP
#   (Optional)
#   If set to a non-empty string, will use it as the URL from which the
#   pip wheel file will be downloaded (instead of building the pip locally).

# export TF_DOCKER_BUILD_CENTRAL_PIP_IS_LOCAL
#   (Optional)
#   If set to a non-empty string, we will treat TF_DOCKER_BUILD_CENTRAL_PIP
#   as a path rather than a url.

export TF_DOCKER_BUILD_IMAGE_NAME=native-mkl-tf
#  (Optional)
#  If set to any non-empty value, will use it as the image of the
#  newly-built image. If not set, the tag prefix tensorflow/tensorflow
#  will be used.

# export TF_DOCKER_BUILD_VERSION:
#   (Optinal)
#   If set to any non-empty value, will use the version (e.g., 0.8.) as the
#   tag prefix of the image. Additional strings, e.g., "-devel-gpu", will be
#   appended to the tag. If not set, the default tag prefix "latest" will be
#   used.

# export TF_DOCKER_BUILD_PORT
#   (Optional)
#   If set to any non-empty and valid port number, will use that port number
#   during basic checks on the newly-built docker image.

# export TF_DOCKER_BUILD_PUSH_CMD
#   (Optional)
#   If set to a valid binary/script path, will call the script with the final
#   tagged image name with an argument, to push the image to a central repo
#   such as gcr.io or Docker Hub.

# export TF_DOCKER_BUILD_PUSH_WITH_CREDENTIALS
#   (Optional)
#   Do not set this along with TF_DOCKER_BUILD_PUSH_CMD. We will push with the
#   direct commands as opposed to a script.

# export TF_DOCKER_USERNAME
#   (Optional)
#   Dockerhub username for pushing a package.

# export TF_DOCKER_EMAIL
#   (Optional)
#   Dockerhub email for pushing a package.

# export TF_DOCKER_PASSWORD
#   (Optional)
#   Dockerhub password for pushing a package.

# export TF_DOCKER_BUILD_PYTHON_VERSION
#   (Optional)
#   Specifies the desired Python version. Defaults to PYTHON2.

# export TF_DOCKER_BUILD_OPTIONS
#   (Optional)
#   Specifies the desired build options. Defaults to OPT.

参考：

MPI教程

tensorflow MPI

build tensorflow

build

install 中文版

pip install mock

REF：

学习课程

more info

conda install for TensorFlow and Intel Distribution for Python upgrade from 2017 to 2018

DNF (Dandified Yum)

Intel® Computer Vision(CV) SDK

Intel's Deep Learning Inference Engine Developer Guide

inference-engine-devguide-introduction

Configuring Model Optimizer for TensorFlow* Prerequisites

Configuring Caffe*

Converting Your TensorFlow* Model

Configuring Model Optimizer for TensorFlow* Prerequisites

What is Intel® DAAL?

应用相关的论文

Pedestrian Detection Using TensorFlow* on Intel® Architecture

tensorflow 监测交通灯

CIFAR-10 分类-tensorflow

构建安装TensorFlow* Serving on Intel® Architecture

Train and Use a TensorFlow* Model on Intel® Architecture

Using the Model Optimizer to Convert TensorFlow* Models

视频 Performance Optimization of Deep Learning Frameworks Caffe* and TensorFlow* for the Intel® Xeon Phi™ Product Family

单节点完整使用教程