快速運行TensorFlow的6種方式

快速運行TensorFlow的6種方式

TensorFlow（http://tensorflow.org）是一個深度學習計算引擎，天然是能夠直接安裝運行的，並且能獲得最佳的性能。可是，考慮到機器學習須要安裝大量的軟件，之間必然會帶來軟件管理和版本兼容性問題，並且在集羣中運行更爲複雜，所以不推薦這種方式。

• 在Linux上安裝TensorFlow，https://www.tensorflow.org/install/install_linux
• 搭建分佈式Tensorflow集羣，http://www.javashuo.com/article/p-bobzteka-r.html
  • 分佈式TensorFlow，中文，http://www.javashuo.com/article/p-rrneniir-gd.html
  • TensorBoard張量圖可視化，https://my.oschina.net/repine/blog/918576
• TensorFlow on Kubernetes架構實踐，https://my.oschina.net/jxcdwangtao/blog/1612667

這裏將介紹幾種能夠在隔離環境中運行TensorFlow的方式，包括：Anaconda、Docker、Jupyter、Kubernetes POD、Kubeflow、Spark DL on MLlib，將更容易管理和擴展到集羣中運行。

• 本文中代碼在：https://github.com/openthings/kubernetes-tools/tree/master/tensorflow

一、Anaconda

經過Anaconda來運行Tensorflow的Python調用代碼，優勢是能夠提供一個相對隔離的Python運行環境，避免與其它的python任務產生版本衝突。雖然也能夠經過VirtualEnv完成相似的任務，但Anaconda自帶包管理功能，用起來更爲方便。

下載和安裝Anaconda:

echo ""
echo "====================================================================="
echo "Downloading Anaconda3 5.1.0 to ~/openthings..."
echo ""

cd ~/openthings
wget -c https://repo.anaconda.com/archive/Anaconda3-5.1.0-Linux-x86_64.sh

echo "====================================================================="
echo "Finished.Latest version at https://repo.continuum.io/archive/"
echo "More practice at https://my.oschina.net/u/2306127/blog"
echo "====================================================================="
echo ""

下載和安裝Miniconda：

echo ""
echo "====================================================================="
echo "Downloading Miniconda3 4.5.1 to ~/openthings..."
echo ""

cd ~/openthings
wget -c https://repo.continuum.io/miniconda/Miniconda3-4.5.1-Linux-x86_64.sh

echo "====================================================================="
echo "Finished.Latest version at https://repo.continuum.io/miniconda/"
echo "More practice at https://my.oschina.net/u/2306127/blog"
echo "====================================================================="
echo ""

安裝TensorFlow on Anaconda：

echo ""
echo "================================================================="
echo "Create conda env for Tensorflow, python 3.6.5 ..."
conda create --yes -n tensorflow pip python=3.6.5
source activate tensorflow
echo ""

echo ""
echo "================================================================="
echo "Install Tensorflow 1.8.0 with GPU support."
pip install --ignore-installed --upgrade \
	https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.8.0-cp36-cp36m-linux_x86_64.whl
echo "================================================================="
echo "More practice at https://my.oschina.net/u/2306127/blog"
echo "More info,Visit https://www.tensorflow.org/install/install_linux "
echo "================================================================="
echo ""

而後運行：

source activate tensorflow

這樣切換Python運行環境，就能夠在Anaconda環境下經過Python執行Tensorflow調用代碼了，固然也能夠安裝Jupyter Notebook之類的組件，而後在瀏覽器裏編寫和運行tensorflow計算任務。

二、Docker

將TensorFlow放到Docker容器中執行，安裝部署都更爲方便，不污染宿主機環境，快速實驗多種版本。

echo ""
echo "================================================================="
echo "Running Tensorflow in Docker with bash shell."
echo "More practice at https://my.oschina.net/u/2306127/blog"
echo "Please Visit https://www.tensorflow.org/install/install_linux "
echo "================================================================="
echo ""

nvidia-docker run -it tensorflow/tensorflow:latest-gpu bash

三、Jupyter

在Docker中安裝Anaconda和Jupyter。

echo ""
echo "================================================================="
echo "Running Tensorflow in Docker with jupyter notebook."
echo "More practice at https://my.oschina.net/u/2306127/blog"
echo "Please Visit https://www.tensorflow.org/install/install_linux "
echo "================================================================="
echo ""

nvidia-docker run -it -p 8888:8888 tensorflow/tensorflow:latest-gpu

而後按照提示的地址，到瀏覽器打開網址 http://localhost:8888，就可使用了。

四、Kubernetes POD

若是使用Kubernetes/Minikube或者OpenShift，能夠將Tensorflow部署到K8S集羣中，參見：

• Kubernetes集成TensorFlow服務，http://www.javashuo.com/article/p-mxgqswxy-mp.html

五、Kubeflow

經過Kubernetes創建TensorFlow機器學習集羣的更佳方式是使用KubeFlow，目前該項目啓動時間不長，還須要進一步完善。使用參考：

• Kubeflow使用指南，http://www.javashuo.com/article/p-nppedcay-hp.html

六、Spark DL on MLlib

因爲TensorFlow只是完成計算功能，經過Spark ML創建深度學習流水線，從而能夠充分利用Spark的分佈式內存和數據處理、數據IO、交互分析等功能，而後進一步將其運行在Kubernetes集羣之中統一調度。

• Apache Spark上的深度學習流水線 ，http://www.javashuo.com/article/p-wafhvhjg-cq.html