DeepHyper Spack Packages

Overview

This repository contains a collection of Spack packages to manage the building of DeepHyper from source. Start by installing Spack on your system:

git clone -c feature.manyFiles=true https://github.com/spack/spack.git
. ./spack/share/spack/setup-env.sh 

Spack is quite recent and evolving quickly. Many packages are regularly updated. Therefore it can sometimes be useful to checkout the develop branch of Spack (git checkout develop) in case of issue to see if a fix has already been provided but not yet released.

Now should have access to the spack command. If you need spack to be loaded for any new session of your shell then you can add . $PATH_TO_SPACK/spack/share/spack/setup-env.sh to your ~/.bashrc (or any other shell configuration). Once Spack is installed and loaded. Download the DeepHyper Spack packages and add them to your Spack installer:

git clone https://github.com/deephyper/deephyper-spack-packages.git
spack repo add deephyper-spack-packages

After adding the deephyper-spack-packages to your Spack installer, create a Spack environment for deephyper:

spack env create deephyper
spack env activate deephyper

You are now ready to install DeepHyper. DeepHyper provides different set of features which not necessary for all users. By default, DeepHyper's installation only brings the basic hyperparameter optimization algorithm with Bayesian optimization and allow parallelism with a centralized scheme (1-manager, n-workers) through basic multi-thread or multi-processing pools.

spack add py-deephyper
spack install

Other features can be installed with the following variants:

spack add deephyper +mpi # MPI support for MPICommEvaluator
spack add deephyper +ray # Ray support for RayEvaluator
spack add deephyper +redis # Redis/RedisJSON/py-redis support for RedisStorage and Distributed Search
spack add deephyper +hps-tl # SDV support for transfer-learning with Bayesian Optimization
spack add deephyper +nas # Tensorflow support for Neural Architecture Search
spack add deephyper +autodeuq # Tensorflow/Tensorflow-Probability/Ray support for Deep Ensemble with Uncertainty Quantitification

# All Features at Once
spack add py-deephyper +autodeuq+hps-tl+mpi+nas+redis

For versions and branches of DeepHyper it is possible to either build from the github repository branches or to build from a release published on pypi.

spack add py-deephyper@master # from master branch on Github
spack add py-deephyper@develop # from develop branch on Github
spack add py-deephyper # from pypi
spack add [email protected] # version 0.4.2 from pypi

The info command can help retrieve more information about the possible variants and versions which can be installed.

spack info py-deephyper

Accessing Shared Site Packages

Sometimes the HPC facility can provide already compiled/installed packages. Spack can reuse these packages by defining a new upstream configuration. After cloning the Spack repository and activating the Spack environment (the command spack should now be available in your shell) add a new upstream configuration in the YAML file $SPACK_ROOT/etc/spack/upstream.yaml ($SPACK_ROOT should be available in your environment after activation of Spack).

At the Argonne Leadership Computing Facility (ALCF), Spack packages are provided at the /soft/spack/root/opt/spack location. Therefore we can edit the upstream.yaml file such as:

upstreams: 
    alcf-spack: 
        install_tree: /soft/spack/root/opt/spack

More information about Chaining Spack Installations.

Spack and Conda

spack add miniconda3

for mac with arm64 use the conda-forge channel!

conda install -c conda-forge ...