Dockerfile

Collection of all my Dockerfile's. The base Dockerfile builds Ubuntu, Miniconda, PyTorch with a default user and a root user. You can access the images on Dockerhub using this link.

Check out my blog post Complete tutorial on building images using Docker that covers everything you need to know to write Dockerfile and run containers using Docker. The base image ubuntu_conda_pytorch is used as an example in the above post and it contains the details of every command used in the above Dockerfile.

ubuntu_conda_pytorch
Customizing base image
- PyTorch CPU only install
- fastai install
License

ubuntu_conda_pytorch

Features

Builds on top of Ubuntu
sudo and a default user
Miniconda setup
PyTorch and torchvision installed using conda
Extra utilities include git, curl, sudo

Build from source Dockerfile

To build the image you can use the following command

> cd ubuntu_conda_pytorch
> docker build -t pytorch:1.9.0 --build-args \
    UBUNTU_VERSION=18.04 \
    USERNAME=default \
    PASSWORD=default \
    MINICONDA_DOWNLOAD_LINK=... \
    MINICONDA_INSTALL_PATH=/home/$USERNAME \
    .

Grab the link of the Linux installer of your choice from the Miniconda document page and use that link as value of MINICONDA_DOWNLOAD_LINK. You can also specify where you want to download Miniconda using MINICONDA_INSTALL_PATH. By default, it is installed in the home directory of the user.

Pull from Dockerhub

The image is hosted at kushaj/pytorch. To pull the image use the following command

> docker pull kushaj/pytorch:1.9.0

Running a container

Use the following command to create a container

> docker run -it --gpus all --name temp kushaj/pytorch:1.9.0

-it will open a terminal connected to the container
--gpus all is used to specify which GPUs to give access to the container
--name temp name of the container

(Usage) Become root

sudo su can be used to become root. The default credentials are

username = default
password = default

(base) default@a96e1cef1e4a:~$ sudo su
[sudo] password for default: 
root@a96e1cef1e4a:/home/default#

(Usage) Nvidia GPU

nvidia-smi can be used to check the status of GPUs. This requires the host machine has Nvidia drivers set up.

(base) default@a96e1cef1e4a:~$ nvidia-smi
Thu Oct  7 02:28:18 2021       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.91.03    Driver Version: 460.91.03    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  Quadro M1200        Off  | 00000000:01:00.0 Off |                  N/A |
| N/A   40C    P0    N/A /  N/A |    356MiB /  4043MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+
(base) default@a96e1cef1e4a:~$

(Usage) PyTorch example

(base) default@7d9b75595a27:~$ python
Python 3.9.7 (default, Sep 16 2021, 13:09:58) 
[GCC 7.5.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True
>>> torch.backends.cudnn.version()
8005
>>> x = torch.tensor([1,2], device='cuda:0')
>>> x
tensor([1, 2], device='cuda:0')

Customizing base image

You can customize ubuntu_conda_pytorch/Dockerfile as per your needs

PyTorch CPU only install

You can modify line 41 of the Dockerfile to install PyTorch. You can head to the PyTorch install page to grab the command to install PyTorch. To install PyTorch for CPU you can change the original command to

> conda install pytorch torchvision cpuonly -c pytorch

fastai install

ubuntu_conda_fastai/Dockerfile contains the code to install fastai in the base repo. You can pull the image from kushaj/fastai.

To run a container use the following command connecting to port 8888 of the container

> docker run -it --gpus all -p 8888:8888 --name temp kushaj/fastai:latest

License

Apache License v2

KushajveerSingh / Dockerfile