From bwHPC Wiki
Jump to: navigation, search

1 Install and use Conda

Conda can be used to easily install missing Python packages on your own into different Python environments with different versions in your own work spaces.

There are two possibilities of using Conda. One to install Conda itself, the other is to load a central installation if possible.

1.1 (A) Installing Conda into a Workspace

We suggest using workspaces for Conda installation. Otherwise $HOME/.conda will be used as default Conda path.

Create workspace for your conda environments:

ws_allocate conda <days>   # e.g. 100, for 100 days
cd $( ws_find conda )

Download script and install conda:

wget -O
bash -b -p $( ws_find conda )/conda

Source variables for conda:

source $( ws_find conda )/conda/etc/profile.d/

1.1.1 Optional Steps

Optional: Add source path in your local ~/.bashrc (edit file):

source $( ws_find conda )/conda/etc/profile.d/

Optional: Update conda (usually base enviroment):

conda update -n base conda

Optional: Removing installation script:


1.2 (B) Use a centrally installed Conda Module

If a Conda Module is provided you can just load it and create environments a workspace.

Load conda module:

ml avail devel/conda   # list Conda modules
ml devel/conda         # load module
# on bwUniCluster:
ml devel/miniconda

Create workspace:

ws_allocate conda <days>   # e.g. 100

The last step defines the newly created workspace as the download and installation path for your environments:

conda config --prepend envs_dirs $( ws_find conda )/conda/envs
conda config --prepend pkgs_dirs $( ws_find conda )/conda/pkgs
conda config --show envs_dirs
conda config --show pkgs_dirs

If you don't specify a new envs_dir Conda will use ~/.conda/envs in your home directory as the default installation path (same applies to pkgs_dirs).

1.3 Install Packages into Environments

You can create python environments and install packages into these environments or create them during install:

conda create -n scipy
conda activate scipy
(scipy) $ conda install scipy

Install packages and create a new environment:

conda create -n scipy scipy
conda activate scipy

Search a special verson:

conda search scipy==1.1.0

Create a Python 2.7 environment:

conda create -n scipy_py27 scipy python=2.7
conda activate scipy_py27

1.4 Activating Environments

In order to use the software in an environment you'll need to activate it first:

conda activate scipy

Deactivate to use different Python or software version:

conda deactivate

Older versions of conda (<4.6) have to use source activate and source deactivate instead.

1.5 List packages and Environments

List packages of current environment:

conda list

List packages in given environment:

conda list -n scipy

List environments:

conda env list

1.6 Use Channels

Add channels to get more software. We suggest to try the following channels:


Search in default and extra channel:

conda search -c intel scipy

You can add channel to your channels, but than you'll search and install automatically from this channel:

conda config --add channels intel
conda config --show channels
conda config --remove channels intel   # remove again

1.7 Use Intel Conda Packages

You can find the full list of Intel Python packages on the Intel web site.

You can install the core Intel Python stack:

conda install -c intel -n intelpython3 intelpython3_core

... with a special Python version:

conda install -c intel -n intelpython-3.6.5 intelpython3_core python=3.6.5

... with a Intel update version:

conda create -c intel -n intelpython-2018.0.3 intelpython3_core==2018.0.3

... or the full Intel Python stack:

conda create -c intel -n intelpython-2018.0.3 intelpython3_full==2018.0.3

... or just some Intel MKL optimized scientific software for the newest Intel 2019 version:

# installs scipy-1.1.0-np115py36_6
conda create -c intel -n scipy-1.1.0-np115py36_6 intelpython3_core=2019

1.8 Create Reproducible Conda Environments

For a more detailed environments documentation refer to the conda documentation.

Create an environment file for re-creation:

conda env export -n scipy-1.1.0-np115py36_6 -f scipy-1.1.0-np115py36_6.yml

Re-create saved environment:

conda env create -f scipy-1.1.0-np115py36_6.yml

Create a file with full URL for re-installation of packages:

conda list --explicit -n scipy-1.1.0-np115py36_6 >scipy-1.1.0-np115py36_6.txt

Install requirements file into environment:

conda create --name scipy-1.1.0 --file scipy-1.1.0-np115py36_6.txt

The first backup option is from the conda-env command and tries to reproduce the environment by name and version. The second option comes from the conda command itself and specifies the location of the file, as well. You can install the identical packages into a newly created environment. Please verify the architecture first.

To clone an existing environment:

conda create --name scipy-1.1.0 --clone scipy-1.1.0-np115py36_6

1.8.1 Local channels and backup Conda packages

Usually packages are cached in your Conda directory inside pkgs/ unless you run conda clean. Otherwise the environment will be reproduced from the channels' packages. If you want to be independent of other channels you can create your own local channel and backup every file you have used for creating your environments.

Install package conda-build:

conda install conda-build

Create local channel directory for linux-64:

mkdir -p $( ws_find conda )/conda/channel/linux-64

Create dependency file list and copy files to channel:

conda list --explicit -n scipy-1.1.0-np115py36_6 >scipy-1.1.0-np115py36_6.txt
for f in $( grep -E '^http|^file' scipy-1.1.0-np115py36_6.txt ); do
    cp $( ws_find conda )/conda/pkgs/$( basename $f ) $( ws_find conda )/conda/channel/linux-64/;

Optional: If packages are missing in the cache download them:

for f in $( grep -E '^http|^file' scipy-1.1.0-np115py36_6.txt ); do
    wget $f -O $( ws_find conda )/conda/channel/linux-64/$( basename $f );

Initialize channel:

conda index $( ws_find conda )/conda/channel/

Add channel to the channels list:

conda config --add channels file://$( ws_find conda )/conda/channel/

Alternative use -c file://$( ws_find conda )/conda/channel/ when installing.

1.8.2 Backup whole Environments

Alternatively you can create a package of your environment and unpack it again when needed.

Install conda-pack:

conda install -c conda-forge conda-pack

Pack activated environment:

conda activate scipy-1.1.0-np115py36_6
(scipy-1.1.0-np115py36_6) $ conda pack
(scipy-1.1.0-np115py36_6) $ conda deactivate

Pack environment located at an explicit path:

conda pack -p $( ws_find conda )/conda/envs/scipy-1.1.0-np115py36_6

The easiest way is to unpack the package into an existing Conda installation.

Just create a directory and unpack the package:

mkdir -p external_conda_path/envs/scipy-1.1.0-np115py36_6
tar -xf scipy-1.1.0-np115py36_6.tar.gz -C external_conda_path/envs/scipy-1.1.0-np115py36_6
conda activate scipy-1.1.0-np115py36_6
# Cleanup prefixes from in the active environment
(scipy-1.1.0-np115py36_6) $ conda-unpack
(scipy-1.1.0-np115py36_6) $ conda deactivate

1.9 Using Singularity container

Using Singularity Containers can create more robust software environments.

Build the container on your local machine!

This is Singularity recipe example for a CentOS image with a Conda environment:

cat << EOF >scipy-1.1.0-np115py36_6.def
BootStrap: yum
OSVersion: 7
Include: yum

# If you want the updates (available at the bootstrap date) to be installed
# inside the container during the bootstrap instead of the General Availability
# point release (7.x) then uncomment the following line

    echo "This is what happens when you run the container..."
    source /conda/etc/profile.d/
    conda activate scipy-1.1.0-np115py36_6
    python --version

    echo "This is what happens when you start the container..."
    source /conda/etc/profile.d/
    conda activate scipy-1.1.0-np115py36_6
    python --version

    echo "Hello from inside the container"
    yum -y install vim wget
    wget -O
    bash -b -p conda
    source /conda/etc/profile.d/
    conda update -y -n base conda
    conda create -y -c intel -n scipy-1.1.0-np115py36_6 intelpython3_core=2019
    rm -f

    source /conda/etc/profile.d/
    conda activate scipy-1.1.0-np115py36_6
    python --version

Build container (on local machine):

singularity build np115py36_6.simg np115py36_6.def

Copy the container on the cluster and start it:

singularity run np115py36_6.simg

See Singularity documentation for more information on containers.

1.10 Versioning

Please keep in mind that modifying, updating and installing new packages into existing environments can modify the outcome of your results. We strongly encourage researchers to creating new environments (or cloning) before installing or updating packages. Consider using meaningful names for your environments using version numbers and dependencies.

Constraint Specification
exact version scipy==1.1.0
fuzzy version scipy=1.1
greater equal "scipy>=1.1"

For more information see cheat sheet below.


conda create -c intel -n scipy-1.1.0 scipy==1.1.0=np115py36_6

1.10.1 Pinning

Pin versions if you don't want them to be updated accidentally (see documentation).


echo 'scipy==1.1.0=np115py36_6' >> $( ws_find conda )/conda/envs/scipy-1.1.0-np115py36_6/conda-meta/pinned

You can easily pin your whole environment:

conda list -n scipy-1.1.0-np115py36_6 --export >$( ws_find conda )/conda/envs/scipy-1.1.0-np115py36_6/conda-meta/pinned

1.10.2 Deleting environments


conda env remove -n scipy-1.1.0-np115py36_6 --all

1.11 Cheat Sheet

Conda official cheat sheet

1.12 PDF Document