伺服器上ssh實作
寫在前面：這隻是我在伺服器上的環境實作的，僅供參考。要根據自己系統的環境做出修改。
==github源碼（https://github.com/mahyarnajibi/SSH）==
**實作參考（https://blog.csdn.net/qq_14845119/article/details/79105360）**
（https://blog.csdn.net/zziahgf/article/details/72900948）
初始工作：安裝cuda和cudnn還有nccl
因為伺服器上裝好了cuda和cudnn，我選擇了cuda9.0和cudnn7.0。是以直接安裝nccl
從github擷取并安裝
git clone https://github.com/NVIDIA/nccl.git
cd nccl
make clean && make PREFIX=$NCCL_ROOT_DIR install
$NCCL_ROOT_DIR是自己安裝的路徑：比如我的路徑是 /home/lzm/data/nccl/install則為：
make clean && make PREFIX=/home/lzm/data/nccl/install install
等nccl安裝完成
安裝caffe-ssh
1、所有都在conda建立的python虛拟環境下進行如：
conda create -n caffetest（虛拟） python=2.7（不是2.7貌似會報錯） anaconda
conda activate caffetest
2、從github擷取源碼：
git clone --recursive https://github.com/mahyarnajibi/SSH.git
3、進入目錄SSH安裝需要的python子產品：
cd SSH
pip install -r requirements.txt
4、建立臨時環境變量env
（1）把nccl和conda環境寫入env檔案：
dlm-conda activate caffetest
export CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/include
export C_INCLUDE_PATH=$C_INCLUDE_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/include
export LIBRARY_PATH=$LIBRARY_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/lib
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/lib
（2）激活環境變量：
source ./env
5、配置檔案将Makefile.config.example拷貝一份成配置檔案：
cd caffe-ssh
cp Makefile.config.example Makefile.config
修改Makefile.config
（1）改成自己cuda的目錄：
CUDA_DIR := /usr/local/cuda
改成
CUDA_DIR := /usr/local/nvidia/cuda/9.0
（2）去掉的注釋：
#OPENCV_VERSION := 3
改成
OPENCV_VERSION := 3
（3）修改環境路徑
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
改為
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu/hdf5/serial/
6、安裝缺少子產品
conda install -c conda-forge readline=6.2
conda install libgcc
7、編譯
make all -j32
8、編譯pycaffe生成接口
make pycaffe
9、在lib中編譯運作setup.py
cd ../lib/
make
10、用scripts中的腳本下載下傳模型
cd ..
bash scripts/download_ssh_model.sh
bash scripts/ download_imgnet_model.sh
11、運作模型示範
python demo.py
結果如下：
可能出現的問題
（1）
問題：
Unsupported gpu architecture 'compute_20'
解決方案：
https://askubuntu.com/questions/960238/nvcc-fatal-unsupported-gpu-architecture-compute-20
即去掉Makefile.config 中兩行：
CUDA_ARCH := -gencode arch=compute_20,code=sm_20 \
-gencode arch=compute_20,code=sm_21 \
-gencode arch=compute_30,code=sm_30 \
-gencode arch=compute_35,code=sm_35 \
-gencode arch=compute_50,code=sm_50 \
-gencode arch=compute_50,code=compute_50
改為：
CUDA_ARCH := -gencode arch=compute_50,code=sm_50 \
-gencode arch=compute_52,code=sm_52 \
-gencode arch=compute_60,code=sm_60 \
-gencode arch=compute_62,code=sm_62 \
-gencode arch=compute_61,code=compute_61
（2）
awk: symbol lookup error: /home/lzm/.conda/envs/lzm2/lib/libreadline.so.6: undefined symbol: PC
https://github.com/conda-forge/rpy2-feedstock/issues/1
https://github.com/bioconda/bioconda-recipes/issues/5350
即 run
conda install -c conda-forge readline = 6.2
（3）
./include/caffe/util/hdf5.hpp:6:18: fatal error: hdf5.h: no such file or directory
https://github.com/BVLC/caffe/issues/2690
https://github.com/NVIDIA/DIGITS/issues/156
即Makefile.config 拿兩行改掉：
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
改為
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu/hdf5/serial/
（4）
./include/caffe/util/nccl.hpp:5:18: fatal error: nccl.h: No such file or directory
建立檔案為env
将伺服器已經安裝的nccl路徑配置到env：
export CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/include
export C_INCLUDE_PATH=$C_INCLUDE_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/include
export LIBRARY_PATH=$LIBRARY_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/lib
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/lzm/data/caffe/caffe1.0_nccl/nccl/install/lib
每次要用的時候都激活環境：
source ./env
（5）
.build_release/lib/libcaffe.so: undefined reference to `cv::imdecode
解決方案：https://github.com/BVLC/caffe/issues/4621
把Makefile.config 中 OPENCV_VERSION = 3的注釋去掉即可
（6）
/caffe/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.21' not found (required by caffe-ssh/python/caffe/_caffe.so)
解決方案：https://github.com/BVLC/caffe/issues/4953
conda install libgcc
PS:以上問題也是自己經過很久的搜尋排查得出來的，不要怕麻煩，要善于搜尋引擎，一切水到渠成