[PYTHON] [0] TensorFlow-GPU environment construction built with Anaconda on Ubuntu

goal

From the place where Ubuntu Desktop / Server is installed, build an environment that can perform GPU calculation by TensorFlow with Anaconda.

Target person

Deep learning engineer who is unfamiliar with Linux and does not understand virtual environments such as Anaconda

equipment

OS: Ubuntu 18.04LTS Desktop/Server

Motivation for this article

I myself wanted to do deep learning on Ubuntu Server, so I built an environment to run the GPU version of TensorFlow, but I was overwhelmed by the number of points to collect information and pay attention to. Therefore, I hope to reduce the pain of connecting knowledge by going back and forth between various pages and by providing a set of information for people with the same aspirations as a memorandum of their own. Therefore, I created a document that explained my knowledge and points of reflection in the process of building the environment.

Document structure

We divided the environment construction into four conditions. I will explain one condition in one article so as not to be complicated. The documentation, including this chapter, is a five-part document (I may also write a sequel to Jupyter Notebook).

This section describes ** 4 conditions ** for building an environment. Clear from top to bottom to run the GPU version of TensorFlow (TensorFlow-GPU).

1. Equipped with a relatively new GPU (manufactured by nVidia) that has the computing power to perform deep learning.
2. The appropriate driver for the GPU is installed and ready for use.
3. An environment has been created to store tensorflow-gpu, GPU drivers (CUDA), and deep learning libraries (cuDNN).
4. Python and each of the conditions 2. must be version compatible

The table of contents is set up according to each condition. You can clear the conditions by working from the top.

[1] TensorFlow-GPU environment built with Anaconda on Ubuntu-GPU spec confirmation-

TensorFlow requires a certain amount of computing power from the GPU. First, let's check in the first article whether the conditions are met.

[2] TensorFlow-GPU environment built with Anaconda on Ubuntu ~ nVidia driver installation ~

In order to use the GPU of the server for calculation, it is necessary to install a dedicated driver in the OS. Since it is a barrier, I will explain it in detail in the second article.

[3] TensorFlow-GPU environment construction built with Anaconda on Ubuntu-Anaconda virtual environment creation-

In addition, there are libraries needed to use the GPU for deep learning. In order to create a foundational environment to accept them, in the third article we will build a virtual environment (described in the article) with Anaconda.

[4] TensorFlow-GPU environment construction built with Anaconda on Ubuntu

~ CUDA, cuDNN installation ~ Finally, in the fourth article, we will install the library on a virtual environment. At this time, if the versions are not compatible, it will not work properly, so I will explain it in one article.

*** First of all [1] TensorFlow-GPU environment built with Anaconda on Ubuntu-GPU spec confirmation- It starts from. *** ***