Deep Learning 有幾種常用的 Tools
* Caffe based on C, but with Python interface: UC Berkeley 開發。Google, CEVA 使用。Mainly for image.
* Torch based on Lua (ANSI C): NYU 開發。Facebook AI and Google use it.
* Theano and Pylearn based on Python : University of Toronto
* TensorFlow: Google
* SystemML: IBM
本文主要 focus 在如何設定 Caffe.
Caffe 幾個特點:
底層是 C (and C++?) 所以速度很快。不過可以用 python (even ipython) 做成 interface. 所以用起來容易。
Protobuf: from Google, for easy and fast data structure.
LMDB: from SAMS, for lightning database
HDF5: from xxx, for database.
OS (Ubuntu 14.04 64bit)
我用 Microsoft surface pro 3 with windows 10. 不過為了方便 try-and-error, 我用 VMware workstation player 12 with guest OS: Ubuntu 14.04 (64-bit)
Install Ubuntu on VMware workstation player 12
* First install VMware workstation player 12 (free)
* Ubuntu 部份:
update Ubuntu’s update --> Yes.
sudo apt-get install tcsh --> No need, just use bash
sudo apt-get install unit-tweak-tool (to increase terminal font size) --> No need, use system setting: displays: scale 1.5 instead!
sudo apt-get install samba (1. edit /etc/smb-conf; 2. smbpasswd -a alu; 3. smb start) --> No need.
-> update Ubuntu 之後，重新 reboot 會有 could not apply the stored configuration for monitors
Solution:
Once logged in change your directory to folder in your home directory.
Rename file to . In here, we are keeping a backup file just to play safe.
Ubuntu Installation
General dependencies
sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler sudo apt-get install --no-install-recommends libboost-all-devCUDA: download from Nvidia (only 64-bit). 如果沒有 GPU 可省略。
BLAS: 使用 ATLAS :
sudo apt-get install libatlas-base-dev
Remaining dependencies 14.04:
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
Python 部份最好後處理
Caffe Installation
Download from Github
* cp Makefile.config.sample Makefile.config
* Edit Makefile.config
Set 1: CPU only, no python layer link. 不支持 pycaffe
make all; make test; make runtest ---> OK (只要先不 install python)
Python (for pycaffe interface)
sudo apt-get install python-dev ----> NONONONO!!!!
不過問題是沒有 numpy. 所以需要 install pip (python package management) and maybe not virtuenv (for coexist of python 2.7 and python 3.x) --> NO!
—> 最直接是只用 python 2.7 且 install Anaconda Python (similar to Enthought python), 不要做 sudo apt-get install python-dev!
Linux Anaconda Installation
- Download the installer.
- After downloading the installer, in your terminal window execute for Python 2.7.10
bash Anaconda2-2.4.0-Linux-x86_64.sh
Anaconda2-2.4.0 是 2.7.10
Anaconda-2.0.0 是 2.7.6 (OK, but need conda upgrade xxx)
bash Ana… .sh to install anaconda python
Option A: 2.4.0 --> (python 2.7.10) ====> NOT GOOD !!!
2.4.0 (python 2.7.10) in /home/alu/anaconda2
Back to caffe and uncomment python part --> make clean; make all; make test OK; but make runtest with error:
libhdf5_hl.so.10 not found.
==> Option B: 2.0.0 --> (python 2.7.6) in /home/alu/anaconda ====> GOOD
No error
> condo update conda
> pip install protobuf
> make pycaffe
How to Use or Verify Python?
Use ipython notebook (in caffe-master directory)
Install boost for python!?? no
Error when running ipython notebook
No module name google.protobuf.internal
> pip install protobuf
> make all; make test ok. but make runtest failed. no libhdf5_hl.so.10 not found. 只有在 /usr/lib/x86_64-linux-gnu/libhdf5_hl.so.7 (這是在 general dependencies 時 install libhdf5-serial-dev 時 Ubuntu 的 default python 是 2.7.6 (?). 但 Anaconda python 是 python 2.7.10.
Ubuntu built-in python lib seems to be out-of-date. Solution:
- install Anaconda python package, and check the location of libhdf5_hl.so.10 一般是在 ~/anaconda/lib
- export LD_LIBRARY_PATH=“$LD_LIBRARY_PATH:~/anaconda/lib” —> work now.
- 需要把 python search path 也設到 .bashrc (anaconda install 自動加入) or .tcshrc; 以及 LD_LIBRARY_PATH
- 檢討是否還是用 equival 來 install 不同版本的 python.
- 重新整理 .bashrc, .bash_aliases, and .tcshrc —> add python search path and LD_LIBRARY_PATH
—> 不過 LD_LIBRARY_PATH 的設定後反而後讓 make all failed! 所以只能在已 compile 之後才加上。
Set 2: CPU only, with python link. has problem!!
