MXNet安装
安装环境：RedHat7
NVIDIA M40
1.下载
git clone --recursive https://github.com/dmlc/mxnet
注意：参数--recursive会更新子模块
| name | description |
|---|---|
| dmlc-core | Data I/O for filesystems such as HDFS and Amazon S3, with job launchers for Yarn, MPI, ... |
| ps-lite | The parameter server framework for asynchronous key-value push and pull |
| mshadow | A lightweight CPU/GPU Matrix/Tensor Template Library. |
依赖：g++ >=4.8 or clang; BLAS library; CUDA Toolkit >=v7.0 for nvidia GPUs; CUDNN; opencv for image augmentation; Atlas
2.编译共享库 libmxnet.so
cd mxnet && make -j16 (并发线程数根据实际环境设置)
问题：
1./usr/bin/ld: cannot find -lcblas
make/config.mk
需要修改ADD_LDFLAGS = -L/usr/lib64/atlas -L/usr/include
mshadow/make/mshadow.mk
需要修改56行 -lcblas => -lsatlas or -ltatlas
2.hdfs.h no such file
原因在于dmlc-core/make/dmlc.mk,设置好HADOOP_HDFS_HOME环境变量,同时，LD_LIBRARY_PATH中要带上libhdfs.so的目录如:$HADOOP_HDFS_HOME/lib/native/
3.cuda.h no such file 如果选择支持GPU，即使用CUDA，需要修改make/config.mk
USE_CUDA_PATH=/usr/local/cuda (安装CUDA的路径)
4.cblas.h no such file
yum install atlas-devel
5.No package 'opencv' found
yum install opencv-devel
3.Python Package Installation
依赖: Python >= 2.7 and numpy
cd python; python setup.py install