Getting started with elektronn3

This guide describes how you can set up an environment for using elektronn3. There are a lot of steps that you could do differently. We won't list all the different options but just explain one way to get it running.

Installing Ananconda

SSH to the server you want to run elektronn3 on and optionally open screen so you can leave the SSH session at any time.

Example:

$ ssh YOUR_USERNAME@example.com
$ screen -rd || screen

Download Miniconda from https://docs.conda.io/en/latest/miniconda.html (or download Anaconda).

Example:

$ wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

Install it to ~/anaconda by running

$ bash Miniconda3-latest-Linux-x86_64.sh

following the assistant and keeping all defaults except the install location, which you should switch to YOUR_HOME_DIRECTORY/anaconda.

Configuring conda

After the installation is complete, create the file ~/.condarc and put the following lines into it:

channels:
  - pytorch
  - conda-forge
  - defaults
pinned_packages:
  - python >=3.6
  - cudatoolkit =9.*  # Pin to a version that your GPU driver supports
  - blas >1.0  # Prevent conda from switching to mkl-blas 1.0
  - pytorch >=1.1.0

The restrictive pinned_packages configuration was at least necessary for me because conda update sometimes does weird things and downgrades essential packages. pinned_packages may not be required on your system.

Creating a conda environment for elektronn3

Now you can create a new conda environment for elektronn3.

$ cd
$ git clone https://github.com/ELEKTRONN/elektronn3
$ conda env create -f elektronn3/environment.yml

This should create a new conda environment called "elektronn3" with all the necessary packages for using it.

Note: If you use other conda envs for different projects, the .condarc above may cause conflicts due to its channel order and pinned_packges restrictions. If it causes problems, you can now move it to your dedicated env to limit its effects to this env:

mv ~/.condarc ~/anaconda/envs/elektronn3/

Note that elektronn3 itself is not yet installed because there is no conda package available for it yet.

Activating the new environment

From now on, always make sure the new environment is active by running

$ conda activate elektronn3

You will see an indicator in your shell that tells you you're in the right env.

Note: The environment activation only lasts for the current shell session. You will need to conda activate elektronn3 in every shell session in which you want to use elektronn3.

Installing the elektronn3 package

To install elektronn3 itself, you will have to use pip:

$ pip install -e --no-deps ~/elektronn3

Explanation of the flags above:

• --no-deps prevents accidental installation of dependencies (this can sometimes happen because pip has problems with dependency resolution)

• -e makes pip install it in "editable" mode. This means that you can (and should) update your local elektronn3 package by navigating to its directory and pulling the latest changes:

  $ cd ~/elektronn3
  $ git pull
  

Next steps

See https://github.com/ELEKTRONN/elektronn3/blob/master/README.md#training