Ubuntu16.04配置各类深度学习框架方法及常见问题

这里主要是GPU版本的框架安装

安装前提：

如果装cpu版本pytorch没有问题(caffe我没有试过)

安装gpu版本需提前安装好nvidia显卡驱动，cuda和cuDNN

其中要保证cuda与nvidia-driver之间版本匹配，cuda与cuDnn版本匹配，cuda与深度学习框架（torch,tensorflow等）版本不配，否则都有可能时出现问题

【1】安装显卡驱动

具体在我的其他博客有写，https://blog.csdn.net/Cindy_lxy/article/details/89438123

【2】安装cuda

下载地址： https://developer.nvidia.com/cuda-toolkit-archive

下载runfile(local)文件

然后直接sh cuda.run

注意：不要安装graph driver,选择这个会重新安装显卡驱动

测试：nvcc -V

如果发现不存在该命令，则需要配置环境变量。（注：不需要安装apt get install nvidia-cuda-toolkit）

配置方法：
gedit ~/.bashrc

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/lib64:/usr/local/cuda/extras/CUPTI/lib64

然后重启一下计算机。nvcc -V就可以看见啦

【3】安装cudnn

下载地址：https://developer.nvidia.com/rdp/cudnn-download

下载  cudnn Library for Linux

然后执行以下命令：

cp cuda/include/cudnn.h /usr/local/cuda/include
cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

测试：

cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2

一、pytorch

简单，明了

https://pytorch.org/get-started/locally/

选一下命令行一复制ok

测试gpu

import torch
torch.cuda.is_available()

二、caffe2

现在caffe2已经集成在pytorch中了

首先安装caffe2需要的环境

https://caffe2.ai/docs/getting-started.html?platform=ubuntu&configuration=compile

Building for source-install dependence

然后安装pytorch就可以了

测试方法：

# To check if Caffe2 build was successful
python -c 'from caffe2.python import core' 2>/dev/null && echo "Success" || echo "Failure"

# To check if Caffe2 GPU build was successful
# This must print a number > 0 in order to use Detectron
python -c 'from caffe2.python import workspace; print(workspace.NumCudaDevices())'

2.编译caffe

在caffe的根目录下：

mkdir build

cd build

cmake ..

make

测试是否安装成功：利用mnist数据集测试caffe是否正确安装

cd ~/caffe/

./data/mnist/get_mnist.sh

./examples/mnist/create_mnist.sh

./examples/mnist/train_lenet.sh

三、tensorflow

pip install tensorflow-gpu==1.11.0

1.注意：如果numpy版本为1.17.0会导致import tensorflow报错

(tensorflow) lxy@mat:~/python/tensorflow/bin$ python
Python 3.5.2 (default, Nov 12 2018, 13:43:14) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
/home/lxy/python/tensorflow/lib/python3.5/site-packages/tensorflow/python/framework/dtypes.py:523: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint8 = np.dtype([("qint8", np.int8, 1)])
/home/lxy/python/tensorflow/lib/python3.5/site-packages/tensorflow/python/framework/dtypes.py:524: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
将版本降为1.16.0即可。

2.如果报这类错误：ImportError: libcusolver.so.8.0: cannot open shared object file: No such file or directory

一般是由于cuda和tensorflow版本不匹配，.8.0就是需要cuda8，解决方法：下载对应版本cuda或者上网查找cuda和tensorflow的版本对应要求。

测试GPU资源是否可用：

tf.test.is_gpu_available()