ubuntu 16.04 安装MXNet GPU版本

安装MXNet for Ubuntu有两种方式。

 

方式一：安装预编译文件

pip install mxnet-cu80

 

方式二：编译源代码

安装nvidia显卡驱动和cuda/cudnn，请参考一下内容 

http://blog.****.net/chenhaifeng2016/article/details/68957732

 

下载mxnet

cd /usr/local/src

git clone https://github.com/dmlc/mxnet.git --recursive

cd mxnet

cp make/config.mk .

vim config.mk

 

cd setup-utils

编辑install-mxnet-ubuntu-python.sh，由于默认不安装在$HOME/mxnet下面，所以需要更改MXNET_HOME

bash install-mxnet-ubuntu-python.sh

 

安装脚本源代码

#!/usr/bin/env bash
######################################################################
# This script installs MXNet for Python along with all required dependencies on a Ubuntu Machine.
# Tested on Ubuntu 14.0 + distro.
######################################################################
set -e

MXNET_HOME="/usr/local/src/mxnet/"
echo "MXNet root folder: $MXNET_HOME"

echo "Installing build-essential, libatlas-base-dev, libopencv-dev, pip, graphviz ..."
sudo apt-get update
sudo apt-get install -y build-essential libatlas-base-dev libopencv-dev graphviz

echo "Building MXNet core. This can take few minutes..."
cd "$MXNET_HOME"
cp make/config.mk .
make -j$(nproc)

echo "Installing Numpy..."
sudo apt-get install python-numpy

echo "Installing Python setuptools..."
sudo apt-get install -y python-setuptools python-pip

echo "Installing Python package for MXNet..."
cd python; sudo python setup.py install

echo "Adding MXNet path to your ~/.bashrc file"
echo "export PYTHONPATH=$MXNET_HOME/python:$PYTHONPATH" >> ~/.bashrc
source ~/.bashrc

echo "Install Graphviz for plotting MXNet network graph..."
sudo pip install graphviz

echo "Installing Jupyter notebook..."
sudo pip install jupyter

echo "Done! MXNet for Python installation is complete. Go ahead and explore MXNet with Python :-)"

 

 

运行example

cd example/image-classification

这里有个坑，运行测试代码前请先安装pip install requests

python train_mnist.py --network mlp

 

使用gpu加速

python train_mnist.py --network mlp --gpus 0

参考资料

http://mxnet.io/get_started/ubuntu_setup.html

转自：https://blog.****.net/chenhaifeng2016/article/details/67634763