Caffe (機械学習)のインストール：CentOS 7.2

Caffe (機械学習ライブラリ）のインストールを CentOS 7.2 搭載サーバに行った。

サーバは以下の二種類
◯計算サーバ１
Huawei RH5885H V3
CPU ：Intel Xeon E7-4890 v2 @ 2.80GHz x 4 socket
Memory ：2.0TB (32GB LRDIMM x 64 DIMMs)
GPU : NVIDIA Tesla K40m x 4
HDD ：2.5-inch 300GB SAS 15000rpm HDD x 2
SSD : ES3000 2.4TB x 2 + 2.5-inch 800GB SSD (Intel DC S3500) x 8
RAID ：RAID-0/1/10/5/50/6/60 1GB Cache with Power Protection
NIC ：On Board 1GE x 4 port interface card
I/O Box ：6 Slot Riser Card x 2、Hot-Plugged Riser Card x 1
PSU ：2000W Platinum AC Power Supply Unit x 2
Rail ：4U Slide Rail with Cable Management Arm
CUDA : 7.5
OS : CentOS 7.2

◯計算サーバ２
CPU : Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz x 1
Memory : 32GB
GPU : NVIDIA GeForce GTX TITAN x 2
CUDA : 7.5
OS : CentOS 7.2

１：様々なライブラリのインストール
> yum install epee-release
> yum install gflags-devel glog-devel lmdb-devel
> yum install protobuf-devel leveldb-devel snappy-devel opencv-devel boost-devel hdf5-devel
> yum install openblas*
> yum install atlas*

２：Caffe と cuDNN のダウンロード
Caffe:
http://caffe.berkeleyvision.org
https://github.com/BVLC/caffe/archive/master.zip

NVIDIA cuDNN
https://developer.nvidia.com/cudnn

３：cuDNN のインストール
tar xvzf cudnn-7.0-linux-x64-v4.0-rc.tgz

以下の５つのファイルを cuda のフォルダ以下にコピー
cuda/include/cudnn.h
cuda/lib64/libcudnn.so
cuda/lib64/libcudnn.so.4
cuda/lib64/libcudnn.so.4.0.4
cuda/lib64/libcudnn_static.a

/sbin/ldconfig の再実行

４：Caffe の make と実行
unzip master.zip
cd caffe-master
cp Makefile.config.example Makefile.config
Makefile.config の編集 http://caffe.berkeleyvision.org/installation.html を参照
make all
make test
make runtest

上記の計算サーバ２では以下のように make した
make -j 6 all
make -j 6 test
make runtest

計算サーバ２では BLAS (Intel MKL と OpenBLAS)によって以下のような数値誤差に違いが生じている。

◯Intel MKL
[----------] Global test environment tear-down
[==========] 1760 tests from 257 test cases ran. (371069 ms total)
[ PASSED ] 1752 tests.
[ FAILED ] 8 tests, listed below:
[ FAILED ] SGDSolverTest/0.TestSnapshotShare, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] SGDSolverTest/0.TestSnapshot, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] AdaGradSolverTest/0.TestSnapshotShare, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] NesterovSolverTest/0.TestSnapshotShare, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] NesterovSolverTest/0.TestSnapshot, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] AdaDeltaSolverTest/0.TestSnapshotShare, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] AdamSolverTest/0.TestSnapshotShare, where TypeParam = caffe::CPUDevice<float>
[ FAILED ] RMSPropSolverTest/0.TestSnapshotShare, where TypeParam = caffe::CPUDevice<float>

8 FAILED TESTS
make: *** [runtest] エラー 1

real 6m13.390s
user 2m44.846s
sys 1m31.608s


◯OpenBLAS
[----------] Global test environment tear-down
[==========] 1760 tests from 257 test cases ran. (375194 ms total)
[ PASSED ] 1760 tests.

real 6m17.486s
user 2m45.805s
sys 1m37.964s