原文：https://makeoptim.com/deep-learning/tensorflow-gpu-on-ubuntu

• 介绍
• 所需软件
• 安装前
  • GCC
  • NVIDIA package repositories
• NVIDIA machine learning
• NVIDIA GPU driver
• CUDA ToolKit and cuDNN
• TensorRT
• Miniconda
• 虚拟环境
• 安装 TensorFlow
• 安装 JupyterLab 和 matplotlib
• 在 JupyterLab 中运行 TensorFlow
• 延伸阅读
• 参考链接

介绍

• Ubuntu 18.04.5 LTS
• GTX 1070
• TensorFlow 2.4.1

所需软件

• NVIDIA® GPU drivers —CUDA® 11.0 需要 450.x 或者更高版本。
• CUDA® Toolkit — TensorFlow 所需 CUDA® 11 (TensorFlow >= 2.4.0)
• cuDNN SDK 8.0.4 cuDNN versions。
• (Optional) TensorRT 6.0 改善延迟和吞吐量，以在某些模型上进行推理。
• Miniconda — 创建虚拟环境。

安装前

GCC

$ gcc --version
Command 'gcc' not found, but can be installed with:
sudo apt install gcc
$ sudo apt install gcc
$ gcc --version
gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

NVIDIA package repositories

$ wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-ubuntu1804.pin
$ sudo mv cuda-ubuntu1804.pin /etc/apt/preferences.d/cuda-repository-pin-600
$ sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub
$ sudo add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/ /"
$ sudo apt-get update

NVIDIA machine learning

$ wget http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb

$ sudo apt install ./nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
$ sudo apt-get update

NVIDIA GPU driver

$ sudo apt-get install --no-install-recommends nvidia-driver-460

注：这里需要使用 460 版本，TensorFlow 官网写的是 450，实测失败。

重启并使用以下命令检查 GPU 是否可见。

$ nvidia-smi
Mon Apr  5 16:17:17 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.32.03    Driver Version: 460.32.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| 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    On   | 00000000:01:00.0  On |                  N/A |
|  0%   48C    P8     9W / 180W |    351MiB /  8111MiB |      1%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A       997      G   /usr/lib/xorg/Xorg                 18MiB |
|    0   N/A  N/A      1145      G   /usr/bin/gnome-shell               53MiB |
|    0   N/A  N/A      1353      G   /usr/lib/xorg/Xorg                108MiB |
|    0   N/A  N/A      1495      G   /usr/bin/gnome-shell               83MiB |
|    0   N/A  N/A      1862      G   ...AAAAAAAAA= --shared-files       82MiB |
+-----------------------------------------------------------------------------+

CUDA ToolKit and cuDNN

$ wget https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/libnvinfer7_7.1.3-1+cuda11.0_amd64.deb
$ sudo apt install ./libnvinfer7_7.1.3-1+cuda11.0_amd64.deb
$ sudo apt-get update

# Install development and runtime libraries (~4GB)
$ sudo apt-get install --no-install-recommends \
    cuda-11-0 \
    libcudnn8=8.0.4.30-1+cuda11.0  \
    libcudnn8-dev=8.0.4.30-1+cuda11.0

TensorRT

$ sudo apt-get install -y --no-install-recommends libnvinfer7=7.1.3-1+cuda11.0 \
    libnvinfer-dev=7.1.3-1+cuda11.0 \
    libnvinfer-plugin7=7.1.3-1+cuda11.0

Miniconda

从 https://docs.conda.io/en/latest/miniconda.html 下载 Python 3.8 安装脚本。

image

增加可执行权限

$ chmod +x Miniconda3-latest-Linux-x86_64.sh

执行安装脚本

$ ./Miniconda3-latest-Linux-x86_64.sh

重启终端，激活 conda。

虚拟环境

创建一个名称为 tensorflow 的虚拟环境。

$ conda create -n tensorflow python=3.8.5
$ conda activate tensorflow

安装 TensorFlow

$ pip install tensorflow==2.4.1

验证安装

$ python -c "import tensorflow as tf;print('Num GPUs Available: ', len(tf.config.list_physical_devices('GPU')))"
2021-04-05 16:20:00.426536: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0
2021-04-05 16:20:01.170305: I tensorflow/compiler/jit/xla_cpu_device.cc:41] Not creating XLA devices, tf_xla_enable_xla_devices not set
2021-04-05 16:20:01.170830: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcuda.so.1
2021-04-05 16:20:01.198917: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:941] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-04-05 16:20:01.199497: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1720] Found device 0 with properties:
pciBusID: 0000:01:00.0 name: GeForce GTX 1070 computeCapability: 6.1
coreClock: 1.7845GHz coreCount: 15 deviceMemorySize: 7.92GiB deviceMemoryBandwidth: 238.66GiB/s
2021-04-05 16:20:01.199519: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0
2021-04-05 16:20:01.201250: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcublas.so.11
2021-04-05 16:20:01.201278: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcublasLt.so.11
2021-04-05 16:20:01.201995: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcufft.so.10
2021-04-05 16:20:01.202159: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcurand.so.10
2021-04-05 16:20:01.203993: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcusolver.so.10
2021-04-05 16:20:01.204412: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcusparse.so.11
2021-04-05 16:20:01.204499: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudnn.so.8
2021-04-05 16:20:01.204566: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:941] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-04-05 16:20:01.204897: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:941] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-04-05 16:20:01.205168: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1862] Adding visible gpu devices: 0
Num GPUs Available:  1

安装 JupyterLab 和 matplotlib

$ pip install jupyterlab matplotlib

在 JupyterLab 中运行 TensorFlow

$ jupyter lab

JupyterLab 将自动在浏览器打开。

从 https://www.tensorflow.org/tutorials/images/cnn 下载并导入 CNN notebook。

image

执行 Restart Kernel and Run All Cells。

image

当训练开始, 检查 GPU 进程，可以看到 ...nvs/tensorflow/bin/python 表示正在使用 GPU 训练模型。

image

$ nvidia-smi
Mon Apr  5 16:36:28 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.32.03    Driver Version: 460.32.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| 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    On   | 00000000:01:00.0  On |                  N/A |
| 23%   54C    P2    72W / 180W |   7896MiB /  8111MiB |     55%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A       997      G   /usr/lib/xorg/Xorg                 18MiB |
|    0   N/A  N/A      1145      G   /usr/bin/gnome-shell               73MiB |
|    0   N/A  N/A      1353      G   /usr/lib/xorg/Xorg                136MiB |
|    0   N/A  N/A      1495      G   /usr/bin/gnome-shell               53MiB |
|    0   N/A  N/A      1862      G   ...AAAAAAAAA= --shared-files       99MiB |
|    0   N/A  N/A      3181      C   ...nvs/tensorflow/bin/python     7507MiB |
+-----------------------------------------------------------------------------+

安装 VSCode

前往官网下载并安装 VSCode。

打开 VSCode 并安装 Python 支持。

image

选择某个文件夹（这里以 ~/tensorflow-notebook/01-hello 为例），新建文件 hello.ipynb。

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
hello.numpy()

使用 VSCode 打开刚才创建的 ~/tensorflow-notebook/01-hello/hello.ipynb，并选择 Python 为创建的虚拟环境。

image

VSCode 运行 TensorFlow

image

小结

至此，开发环境已经搭建完毕。大家可以根据自己的习惯，选择使用命令行、JupyterLab 或者 VSCode 进行开发。

延伸阅读

• Mac 机器学习环境 (TensorFlow, JupyterLab, VSCode)
• Mac M1 机器学习环境 (TensorFlow, JupyterLab, VSCode)
• Win10 机器学习环境 (TensorFlow GPU, JupyterLab, VSCode)

参考链接

• https://www.tensorflow.org/install
• https://github.com/tensorflow/tensorflow/issues/39526
• https://www.tensorflow.org/guide/gpu