Install DeepHyper with pip

DeepHyper installation requires Python>=3.7.

Warning

All packages required for your application need to be installed in the same environment.

DeepHyper is available on PyPI and can be installed with pip on Linux, macOS by downloading pre-built binaries (wheels) from PyPI:

$ # Default set of features (HPS, NAS, AutoDEUQ, Transfer-Learning and LCE Stopper)
$ pip install "deephyper[default]" # <=> "deephyper[hps,nas,autodeuq,hps-tl,jax-cpu]"

$ # Isolated features
$ pip install "deephyper[hps]" # Install Hyperparameter Search.
$ pip install "deephyper[nas]" # Install Neural Architecture Search.
$ pip install "deephyper[autodeuq]" # Install Automated Deep Ensemble with Uncertainty Quantification.
$ pip install "deephyper[hps-tl]" # Install Transfer-Learning for HPS.
$ pip install "deephyper[jax-cpu]" # Install JAX with CPU support for Learning Curve Extrapolation Stopper.
$ pip install "deephyper[jax-cuda]" # Install JAX with GPU (cuda) support for Learning Curve Extrapolation Stopper.
$ pip install "deephyper[automl]" # Install Automated Machine Learning features.

$ # Others
$ pip install "deephyper[analytics]" # Install Analytics tools (for developers).
$ pip install "deephyper[dev]" # Install Developer Stack (tests, documentation, etc...)

In Bayesian optimization, the Mondrian Forest surrogate model can be used. This model provides better uncertainty estimates used in the acquisition function. To install the Mondrian Forest surrogate model, you need to install the modified scikit-garden package from our repository. This package is not available on PyPI but can be installed through pip from the GitHub repository:

$ pip install "scikit-garden @ git+https://github.com/deephyper/scikit-garden@master"

Distributed Computation

DeepHyper supports distributed computation with different backends. MPI demonstrated better scaling capabilities but Ray is more flexible and easier to use on smaller scales. Redis is required for distributed search (i.e., different search instances communicating with each other through a shared database). The following command install the “client” or “python binding” of the corresponding libraries but for MPI and Redis you will also need to install the corresponding libraries on your system prior to the pip ... command.

$ pip install "deephyper[mpi]" # Install MPI features for MPICommEvaluator.
$ pip install "deephyper[ray]" # Install Ray features for RayEvaluator.
$ pip install "deephyper[redis]" # Install Redis Client for RedisStorage with Distributed Search.
$ pip install "deephyper[redis-hiredis]" # Install Redis with Hiredis for better performance.

For Redis we advice to follow the Redis official installation guide to install the client/server features. Then, the RedisJson also needs to be installed by following the Redis JSON official installation guide. To make the installation easier we provide a Spack package redisjson for which installation instruction are detailed at Redis Server & RedisJSON.

For MPI we advice to follow the MPI official installation guide to install the client/server features. But, in many centers an MPI installation will already be provided or it can also be installed through a package manager (e.g., apt or brew).