Skip to content

setting-up-for-ml.md

Latest commit

 

History

History
157 lines (111 loc) · 4.71 KB

setting-up-for-ml.md

File metadata and controls

157 lines (111 loc) · 4.71 KB

Setting up docker for ML

Machine Learning (ML) requires access to serious computing hardware, most notably Graphical Processing Units (GPUs) capable of crunching vast amount of data in order to train complicated deep (learning) neural networks models. NVidia is currently the dominant GPU provider on the market. A large fraction of deep learning models rely upon NVidia GPUs. Before we setup a docker environment that has access to GPUs available on the host machine, lets ensure that the following works. To keep things simple, lets assume that you are using a Linux system.

  1. Nvidia GPU hardware is correctly installed on your machine.
  2. You have downloaded and installed the latest drivers for your GPU/OS combination.
  3. You have also installed the correct version of Nvidia cuda toolkit.
  4. In addition, you have also installed Docker and docker-compose.

Which GPUs are available

Use the following commands to see GPU hardware available on your machine.

$ nvidia-smi

Which version of cuda is installed

Use the following command to check cuda compiler version installed on your machine.

$ nvcc --version

Which version of linux is this system

Check linux version

$ cat /etc/*release

Alternately, you can use lsb_release -d command.

You can install lsb package if not already installed

$ apt-get -y install lsb-core

Nvidia Container Toolkit

Nvidia has provided Nvidia Container Toolkit package that allows GPUs to be made available within the docker container. Install it as follows

$ sudo apt-get install -y nvidia-docker2
$ sudo systemctl restart docker

Setting up Nvidia/cuda container

You're almost there.

Option 1

You can docker pull to get a pre-built container. Check here and here for more information.

$ docker pull nvidia/cuda:11.3.1-cudnn8-runtime-ubuntu20.04

This will require docker login.

Option 2

This is my preferred way of doing things. Create a Dockerfile as follows. The first line is important and it specifies the base image nvidia/cuda image to build from. The rest is standard docker stuff as seen here.

Dockerfile

FROM nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04

ARG DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get -y upgrade

RUN apt-get install -y build-essential python3 python3-pip python-dev sudo

RUN mkdir -p /tmp
COPY requirements.txt /tmp/
RUN pip3 -q install pip --upgrade
RUN pip3 install -r /tmp/requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

RUN groupadd -g 1010 dockeruser
RUN useradd -r -m -g 1010 dockeruser
RUN chown -R dockeruser /home/dockeruser
RUN chmod -R g+rwx /home/dockeruser
USER dockeruser

Since I am interested in PyTorch, my requirements.txt file contains the following.

requirements.txt

torch==1.8.0+cu111
torchvision==0.9.0+cu111
matplotlib
jupyter

Find out the appropriate torch package with the correct cuda support here. This sort of needs to match with the cuda support available on the host and also the nvidia/cuda base image that you use as the base image (as seen in the docker file).

Building your own nvidia/cuda docker container

Build docker container as usual.

$ docker build -t myml .

Run container

You can run docker container as follows

$ docker run -it --gpus=all --ipc=host myml bash

From within you can see if PyTorch is able to access host GPUs as follows.

[Container] $ python3
Python 3.6.9 (default, Jan 26 2021, 15:33:00)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True
>>> torch.cuda.get_device_name(0)
'GeForce GTX TITAN X'
>>> torch.cuda.get_device_name(1)
'GeForce GTX 980'
>>>

Happy hacking!