- Cuda
如果配了Nvidia卡的,可以考虑安装Cuda,这样之后可以用GPU加速。之前写过一篇在Ubuntu 14.04上装Cuda 7.5的文章(Link)。TensorFlow 1.2版本貌似需要Cuda Toolkit 8.0,过程和之前是差不多的。更新driver(如需),然后去Nvidia官网下载Cuda和cuDNN安装即可。具体不再累述。对于大部分N卡,Cuda 8.0需要driver的最低版本为367,所以如果已经够用,在安装cuda的时候保险点的话就不用更新驱动。如果更新驱动后不幸中招,如循环登录或无法进入图形界面等问题,可以到字符终端(CTL+ALT+F1)先尝试清除已有驱动,禁用Nvidia开源驱动nouveau,然后重装驱动。
sudo apt-get remove --purge nvdia*在blacklist.conf中加上: blacklist nouveau blacklist lbm-nouveau options nouveau modeset=0 alias nouveau off alias lbm-nouveau off
sudo apt-get install update
sudo apt-get install dkms build-essential linux-headers-generic
sudo vim /etc/modprobe.d/blacklist.conf
sudo service lightdm stop重启。如果进不了图形界面,就把unity那坨都重装一下,然后再通过sudo service lightdm start启动桌面环境。
sudo add-apt-repository ppa:graphics-drivers/ppa && sudo apt-get update
sudo apt-get install nvidia-375
- Anaconda
Anaconda发行版可以用于创建独立的python开发运行环境。每个环境中的python runtime都是独立的,互不影响。这样就不用担心安装A的时候把B的环境给破坏了。Anaconda最新版本4.4.0。下载链接为:https://www.continuum.io/downloads。安装很方便,以Anaconda for Python 2.7为例:
bash ~/Downloads/Anaconda2-4.4.0-Linux-x86_64.sh然后就可以创建环境,比如创建两个分别为python 2.7和3.5的环境:
conda create --name py35 python=3.5其中py27和py35为环境名,之后用
conda create --name py27 python=2.7
source activate <env name>进入相应的环境。删除环境可以用:
conda remove --name <env name> --all列出现有的环境:
conda env list列出环境中安装的包:
conda list --name=<env name>更多用法请参见:https://conda.io/docs/using/envs.html
进入环境后安装包既可以用conda install也可以用传统的pip install,有时网络不给力的时候可能下载会超时: ReadTimeoutError: HTTPSConnectionPool(host='pypi.python.org', port=443): Read timed out. 如果真的只是因为慢,这里可以用延长timeout时间来解决:
pip --default-timeout=10000 install -U <package name>另外如果在使用过程中碰到下面错误: ValueError: failed to parse CPython 有可能是和用户目录下的本地环境串了。一个方法是打开anaconda2/lib/python2.7/site.py,修改ENABLE_USER_SITE = False。
- TensorFlow
目前最新release版本为1.2.1(1.3还是RC状态)。我们就以v1.2.1为例。最方便的话就是装prebuild版:https://www.tensorflow.org/install/install_linux。如果已经装了Anaconda,先进入环境(假设已经创建了python 2.7的环境,名为py27):
source activate py27如果没有安装Anaconda的话上面这步就省了。之后安装TensorFlow,其中的binary下载链接需要根据python版本,有无GPU信息在https://www.tensorflow.org/install/install_linux#the_url_of_the_tensorflow_python_package中自行选取。如python 3.5,有GPU的情况下就可以用:
pip install --ignore-installed --upgradehttps://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.2.1-cp35-cp35m-linux_x86_64.whl再稍微验证下能否顺利加载:
python -c "import tensorflow as tf;print(tf.__version__);"如果打印出刚装的版本号那就差不多了。
但官方prebuild版没有加入x86并行指令(SSE/AVX/FMA)优化。因此训练的时候会打印类似下面信息:
2017-08-12 20:10:39.973508: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.有个鸵鸟的办法就是将log level提高,眼不见心不烦:
2017-08-12 20:10:39.973536: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-12 20:10:39.973541: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-08-12 20:10:39.973545: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-12 20:10:39.973549: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
export TF_CPP_MIN_LOG_LEVEL=2但这样把其它一些log也过滤了。另一方面,x86的并行加速指令在一些情况下是可以带来几倍的性能提升的。因此我们可以考虑自己编译一个带该优化的版本。先下载源码,然后checkout相应版本分支(如r1.2):
git clone https://github.com/tensorflow/tensorflow参考https://*.com/questions/41293077/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions,安装好编译工具bazel后(https://docs.bazel.build/versions/master/install-ubuntu.html),可以用以下命令编译:
git checkout r1.2
bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda -k //tensorflow/tools/pip_package:build_pip_package如果你编译的时候碰到以下错误:
Loading:这是一个已知问题(https://github.com/tensorflow/tensorflow/pull/11949),解决方法见https://github.com/tensorflow/tensorflow/pull/11949/commits/c5d311eaf8cc6471643b5c43810a1feb19662d6c,目前貌似还没有pick到发布分支,人肉pick下吧,应该就解决了。编译好后用下面命令在指定目录(如~/tmp/)生成whl安装包,然后就和前面一样安装即可。
Loading: 0 packages loaded
ERROR: Skipping '//tensorflow/tools/pip_package:build_pip_package': error loading package 'tensorflow/tools/pip_package': Encountered error while reading extension file 'cuda/build_defs.bzl': no such package '@local_config_cuda//cuda': Traceback (most recent call last):
bazel-bin/tensorflow/tools/pip_package/build_pip_package ~/tmp/如果运行时出现下面错误:
ImportError: Traceback (most recent call last):根据https://*.com/questions/35953210/error-running-basic-tensorflow-example,cd到非tensorflow源码目录即可。
File "tensorflow/python/pywrap_tensorflow.py", line 41, in <module>
from tensorflow.python.pywrap_tensorflow_internal import *
ImportError: No module named pywrap_tensorflow_internal