TensorFlow Install

1) 确认主机支持avx指令
[root@srv1 ~]# cat /proc/cpuinfo | grep avx

2) 安装Python 3.6
[root@srv1 ~]# yum install python3 -y

3) 安装一些软件包
[root@srv1 ~]# yum install python3-devel python3-virtualenv gcc gcc-c++ make -y

4) 安装TensorFlow
[root@srv1 ~]# su - snow
[snow@srv1 ~]$ virtualenv --system-site-packages -p python3 ./venv
Running virtualenv with interpreter /usr/bin/python3
Using base prefix '/usr'
New python executable in /home/snow/venv/bin/python3
Also creating executable in /home/snow/venv/bin/python
Installing setuptools, pip, wheel...done.

[snow@srv1 ~]$ source ./venv/bin/activate
(venv) [snow@srv1 ~]$

(venv) [snow@srv1 ~]$ pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade tensorflow==2.0.0b1

# 在TensorFlow 2.0及以上版本中，如果numpy高于1.17则会出现许多警告。因此需要降级numpy至1.16版本
(venv) [snow@srv1 ~]$ pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade numpy==1.16.0
(venv) [snow@srv1 ~]$ pip3 show numpy
Name: numpy
Version: 1.16.0
Summary: NumPy is the fundamental package for array computing with Python.
Home-page: https://www.numpy.org
Author: Travis E. Oliphant et al.
Author-email: None
License: BSD
Location: /home/cent/venv/lib/python3.6/site-packages
Requires:
Required-by: tensorflow, tb-nightly, Keras-Preprocessing, Keras-Applications, h5py

5) 验证TensorFlow
(venv) [snow@srv1 ~]$ python3 -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
2020-07-26 23:46:23.047533: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2801355000 Hz
2020-07-26 23:46:23.048084: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x5633e28df970 executing computations on platform Host. Devices:
2020-07-26 23:46:23.048148: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): <undefined>, <undefined>
tf.Tensor(545.2815, shape=(), dtype=float32)

(venv) [snow@srv1 ~]$ python3 -c "import tensorflow as tf; hello = tf.constant('Hello, TensorFlow World'); tf.print(hello)"
2020-07-26 23:48:13.047533: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2801355000 Hz
2020-07-26 23:48:13.815812: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x5633f4d64910 executing computations on platform Host. Devices:
2020-07-26 23:48:13.815899: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): <undefined>, <undefined>
Hello, TensorFlow World

1) 安装CUDA 10.1

2) 安装Python 3.8

3) 下载cuDNN(CUDA Deep Neural Network library)
 https://developer.nvidia.com/rdp/cudnn-download

 4) 安装cuDNN
[root@srv1 ~]# tar zxvf cudnn-10.1-linux-x64-v7.6.5.32.tgz
[root@srv1 ~]# cp ./cuda/include/cudnn.h /usr/local/cuda-10.1/include/
[root@srv1 ~]# cp -a ./cuda/lib64/libcudnn* /usr/local/cuda-10.1/lib64/
[root@srv1 ~]# ldconfig
[root@srv1 ~]# echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.1/extras/CUPTI/lib64' >> /etc/profile.d/cuda101.sh
[root@srv1 ~]# source /etc/profile.d/cuda101.sh

5) 安装一些必要的软件
[root@srv1 ~]# yum install centos-release-scl-rh centos-release-scl -y

[root@srv1 ~]# sed -i -e "s/\]$/\]\npriority=10/g" /etc/yum.repos.d/CentOS-SCLo-scl.repo
[root@srv1 ~]# sed -i -e "s/\]$/\]\npriority=10/g" /etc/yum.repos.d/CentOS-SCLo-scl-rh.repo

[root@srv1 ~]# sed -i -e "s/enabled=1/enabled=0/g" /etc/yum.repos.d/CentOS-SCLo-scl.repo
[root@srv1 ~]# sed -i -e "s/enabled=1/enabled=0/g" /etc/yum.repos.d/CentOS-SCLo-scl-rh.repo

[root@srv1 ~]# yum --enablerepo=centos-sclo-rh install rh-python38-python-devel gcc gcc-c++ make -y
[root@srv1 ~]# pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple virtualenv

6) 安装TensorFlow 2.2
[root@srv1 ~]# su - snow
[snow@srv1 ~]$ virtualenv --system-site-packages -p python3 ./venv

[snow@srv1 ~]$ source ./venv/bin/activate
(venv) [snow@srv1 ~]$ pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade tensorflow==2.2.0

7) 验证
(venv) [snow@srv1 ~]$ python3 -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
2020-07-26 23:58:07.037218: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcuda.so.1
......
......

1) 安装Docker-CE
[root@srv1 ~]# yum install yum-utils -y
[root@srv1 ~]# yum-config-manager --add-repo https://download.docker.com/linux/centos/docker-ce.repo
[root@srv1 ~]# yum install docker-ce -y
[root@srv1 ~]# systemctl enable --now docker

[root@srv1 ~]# vim /etc/docker/daemon.json
{
      "registry-mirrors": ["https://3laho3y3.mirror.aliyuncs.com"]
}

[root@srv1 ~]# systemctl restart docker

2) 下载Tensorflow
[root@srv1 ~]# su - snow
[snow@srv1 ~]$ podman pull tensorflow/tensorflow:2.0.0-py3

[snow@srv1 ~]$ podman images
REPOSITORY                        TAG         IMAGE ID       CREATED        SIZE
docker.io/tensorflow/tensorflow   2.0.0-py3   90f5cb97b18f   9 months ago   1.09 GB

3) 运行Tensorflow容器
[snow@srv1 ~]$ podman run --rm tensorflow/tensorflow:2.0.0-py3 \
python -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"

2020-07-27 00:23:15.128039: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2801480000 Hz
......
......

[snow@srv1 ~]$ vim hello_tensorflow.py
import tensorflow as tf
hello = tf.constant('Hello, TensorFlow World!')
tf.print(hello)

[snow@srv1 ~]$ podman run --rm -v $PWD:/tmp -w /tmp tensorflow/tensorflow:2.0.0-py3 python ./hello_tensorflow.py
2020-07-27 00:25:03.110934: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2801480000 Hz
2020-07-27 00:25:03.380641: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x4add050 executing computations on platform Host. Devices:
2020-07-27 00:25:03.936027: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): Host, Default Version
Hello, TensorFlow World!

4) 设置SELinux
[root@srv1 ~]# vim my-python.te
module my-python 1.0;


require {
        type user_home_t;
        type container_t;
        type user_home_dir_t;
        class file { create ioctl open read unlink write };
        class dir { add_name remove_name write };
}


#============= container_t ==============
allow container_t user_home_dir_t:dir { add_name remove_name write };
allow container_t user_home_dir_t:file { create ioctl open read unlink write };
allow container_t user_home_t:file { ioctl open read };

[root@srv1 ~]# checkmodule -m -M -o my-python.mod my-python.te
[root@srv1 ~]# semodule_package --outfile my-python.pp --module my-python.mod
[root@srv1 ~]# semodule -i my-python.pp

5) 安装并运行带有Jupyter Notebook的TensorFlow镜像
[snow@srv1 ~]$ podman pull tensorflow/tensorflow:2.0.0-py3-jupyter

[snow@srv1 ~]$ podman images
REPOSITORY                        TAG                 IMAGE ID       CREATED        SIZE
docker.io/tensorflow/tensorflow   2.0.0-py3-jupyter   c652a4fc8a4f   9 months ago   1.24 GB
docker.io/tensorflow/tensorflow   2.0.0-py3           90f5cb97b18f   9 months ago   1.09 GB

[snow@srv1 ~]$ podman run -dt --name jn -p 8888:8888 tensorflow/tensorflow:2.0.0-py3-jupyter
a20ffeda3160ea59d1efd4182fb96a93bae4df4c764befaadc57c1ee77fcbcd1

[snow@srv1 ~]$ podman exec jn bash -c "jupyter notebook list"
Currently running servers:
http://0.0.0.0:8888/?token=e72eefb6fa80c74d9640ec9db68a25fd68484d527126bdf6 :: /tf

6) 访问Jupyter Notebook
复制"http://0.0.0.0:8888/?token=e72eefb6fa80c74d9640ec9db68a25fd68484d527126bdf6",并将0.0.0.0改成本机IP地址
如:
http://192.168.10.111:8888/?token=e72eefb6fa80c74d9640ec9db68a25fd68484d527126bdf6

1) 主机支持avx并安装好NVIDIA显卡驱动

2) 安装好Docker-CE

3) 安装NVIDIA Container Toolkit

[root@srv1 ~]# curl https://nvidia.github.io/nvidia-docker/centos8/nvidia-docker.repo > /etc/yum.repos.d/nvidia-docker.repo

[root@srv1 ~]# yum install nvidia-container-toolkit -y

4) 设置SELinux
[root@srv1 ~]# vim nvidiasmi.te
module my-python 1.0;


module nvidiasmi 1.0;


require {
        type container_runtime_tmpfs_t;
        type container_t;
        type xserver_misc_device_t;
        class file { open read };
        class chr_file { getattr ioctl open read write };
}


#============= container_t ==============
allow container_t container_runtime_tmpfs_t:file { open read };
allow container_t xserver_misc_device_t:chr_file { getattr ioctl open read write };

[root@srv1 ~]# checkmodule -m -M -o nvidiasmi.mod nvidiasmi.te
[root@srv1 ~]# semodule_package --outfile nvidiasmi.pp --module nvidiasmi.mod
[root@srv1 ~]# semodule -i nvidiasmi.pp

5) 使用nvidia-smi
[root@srv1 ~]# docker run --gpus all --rm nvidia/cuda:11.0-base nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.57       Driver Version: 450.57       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1070    Off  | 00000000:05:00.0 Off |                  N/A |
| 27%   35C    P5    27W / 180W |      0MiB /  8119MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+


+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

root@62ffc368879:/# nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.57       Driver Version: 450.57       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1070    Off  | 00000000:05:00.0 Off |                  N/A |
| 27%   36C    P5    25W / 180W |      0MiB /  8119MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+


+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

root@62ffc368879:/# exit

[root@srv1 ~]# podman images
REPOSITORY              TAG         IMAGE ID       CREATED        SIZE
docker.io/nvidia/cuda   11.0-base   27c1a1745519   4 days ago     125 MB
docker.io/nvidia/cuda   10.0-base   841d44dd4b3c   7 months ago   113 MB

6) 下载CUDA/TensorFlow镜像
[root@srv1 ~]# docker pull nvidia/cuda:10.1-base
[root@srv1 ~]# docker pull tensorflow/tensorflow:2.1.0-gpu-py3

[root@srv1 ~]# docker images
REPOSITORY                        TAG             IMAGE ID       CREATED        SIZE
docker.io/tensorflow/tensorflow   2.1.0-gpu-py3   e2a4af785bdb   6 months ago   4.13 GB
docker.io/nvidia/cuda             10.1-base       3b55548ae91f   7 months ago   109 MB

7) 验证
[root@srv1 ~]# docker run --gpus all --rm nvidia/cuda:10.1-base nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.57       Driver Version: 450.57       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1070    Off  | 00000000:05:00.0 Off |                  N/A |
| 27%   35C    P5    25W / 180W |      0MiB /  8119MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+


+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

8) 验证并运行TensorFlow
[root@srv1 ~]# docker run --gpus all --rm tensorflow/tensorflow:2.1.0-gpu-py3 \
python -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
2020-07-27 00:48:25.131105: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer.so.6
......
......