Caffe (Convolution Architecture For Feature Extraction)做爲深度學習CNN一個很是火的框架,對於初學者來講,搭建Linux下的Caffe平臺是學習深度學習關鍵的一步,其過程也比較繁瑣,回想起當初折騰的那幾天,遂總結一下Ubuntu14.04的配置過程,方便之後新手能在此少走彎路。python
1. 安裝build-essentialslinux
安裝開發所須要的一些基本包git
1: sudo apt-get install build-essential
2. 安裝NVIDIA驅動github
輸入下列命令添加驅動源shell
1: sudo add-apt-repository ppa:xorg-edgers/ppa
2: sudo apt-get update
安裝340版本驅動(具體版本取決於電腦顯卡的型號,詳細可到NVIDIA官網查看)ubuntu
1: sudo apt-get install nvidia-340
安裝完成後,繼續安裝下列包bash
1: sudo apt-get install nvidia-340-uvm
安裝驅動完畢,reboot.app
3. 安裝CUDA 6.5框架
CUDA的Deb包安裝較爲簡單,按照官網流程,事先安裝必要的庫學習
1: sudo apt-get install freeglut3-dev build-essential libx11-dev libxmu-dev libxi-dev libgl1-mesa-glx libglu1-mesa libglu1-mesa-dev
3.1 安裝CUDA
而後經過如下命令獲取Ubuntu 14.04 CUDA相關的repository package
1: $ sudo dpkg -i cuda-repo-ubuntu1404_6.5-14_amd64.deb
2: $ sudo apt-get update
而後開始安裝CUDA Toolkit
1: $ sudo apt-get install cuda
此時須要下載較長時間,網速較慢的中途能夠出去吃個飯~
3.2 環境配置
CUDA安裝完畢後,須要對.bashrc加入一下命令來配置環境
1: export CUDA_HOME=/usr/local/cuda-6.52: export LD_LIBRARY_PATH=${CUDA_HOME}/lib64
1: PATH=${CUDA_HOME}/bin:${PATH}2: export PATH
經過複製SDK samples 到主目錄下,完成整個編譯過程
1: $ cuda-install-samples-6.5.sh ~
2: $ cd ~/NVIDIA_CUDA-6.5_Samples
3: $ make
若是以上過程都成功後,能夠經過運行bin/x86_64/linux/release 下的deviceQuery來驗證一下。若是出現如下信息,則說明驅動以及顯卡安裝成功
1: ./deviceQuery Starting...
2:
3: CUDA Device Query (Runtime API) version (CUDART static linking)
4:
5: Detected 1 CUDA Capable device(s)
6:
7: Device 0: "GeForce GTX 670"
8: CUDA Driver Version / Runtime Version 6.5 / 6.5
9: CUDA Capability Major/Minor version number: 3.0
10: Total amount of global memory: 4095 MBytes (4294246400 bytes)
11: ( 7) Multiprocessors, (192) CUDA Cores/MP: 1344 CUDA Cores
12: GPU Clock rate: 1098 MHz (1.10 GHz)
13: Memory Clock rate: 3105 Mhz
14: Memory Bus Width: 256-bit
15: L2 Cache Size: 524288 bytes
16: Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
17: Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
18: Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
19: Total amount of constant memory: 65536 bytes
20: Total amount of shared memory per block: 49152 bytes
21: Total number of registers available per block: 65536
22: Warp size: 32
23: Maximum number of threads per multiprocessor: 2048
24: Maximum number of threads per block: 1024
25: Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
26: Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
27: Maximum memory pitch: 2147483647 bytes
28: Texture alignment: 512 bytes
29: Concurrent copy and kernel execution: Yes with 1 copy engine(s)
30: Run time limit on kernels: Yes
31: Integrated GPU sharing Host Memory: No
32: Support host page-locked memory mapping: Yes
33: Alignment requirement for Surfaces: Yes
34: Device has ECC support: Disabled
35: Device supports Unified Addressing (UVA): Yes
36: Device PCI Bus ID / PCI location ID: 1 / 0
37: Compute Mode:
38: < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >
39:
40: deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 6.5, CUDA Runtime Version = 6.5, NumDevs = 1, Device0 = GeForce GTX 670
41: Result = PASS
4. 安裝BLAS
Caffe的BLAS能夠有三種選擇,分別爲atlas、mkl以及openBLAS。對於mkl能夠到intel官網下載,解壓完成後又一個install_GUI.sh文件,執行該文件會出現圖形安裝界面,根聽說明一步一步執行便可。
也可對openBLAS源碼進行編譯,不過須要gcc以及gfortran等相關編譯器。我的認爲比較便捷的是atlas,在Caffe官網上有相關的介紹,對於Ubuntu,經過如下命令能夠下載atlas
1: sudo apt-get install libatlas-base-dev
5. 安裝OpenCV
OpenCV庫安裝能夠經過網上寫好的腳本進行下載:https://github.com/jayrambhia/Install-OpenCV
解壓文檔後,進入Ubuntu/2.4 給全部的shell腳本加上可執行權限
1: chmod +x *.sh
而後執行 opencv2_4_9.sh 安裝最新版本,注意,OpenCV 2.4.9不支持gcc-4.9以上的編譯器!!
6. 安裝其餘dependencies
對於Ubuntu 14.04,執行如下命令下載其餘相關依賴庫文件
1: sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libboost-all-dev libhdf5-serial-dev
2: sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev protobuf-compiler
7. 安裝python以及Matlab
首先安裝pip和python –dev
1: sudo apt-get install python-dev python-pip
以及caffe python wrapper所須要的額外包
1: sudo pip install -r /path/to/caffe/python/requirements.txt
Matlab接口須要額外安裝Matlab程序
Last shot --- 編譯Caffe
完成全部的環境配置,終於能夠編譯caffe了,經過官網下載caffe源碼,進入根目錄caffe-master,首先複製一份makefile
1: cp Makefile.config.example Makefile.config
而後修改裏面的內容,主要有:
CPU_ONLY 是否採用cpu模式,不然選擇CUDNN(這裏的CUDNN須要在NVIDIA-CUDNN下載,還有經過email註冊申請才能經過審覈)
BLAS:=atlas(也能夠是open或者mkl)
DEBUG 若是須要debug模式
MATLAB_DIR 若是須要採用matlab 接口
完成配置後,能夠進行編譯了
1: make all -j4
2: make test
3: make runtest
最後若是都能正常,證實caffe裏面全部的例子程序均可以運行了,放心都跑CIFAR、MNIST以及ImageNet吧~~