Installation¶

TL;DR Use conda install not pip and for real work, run fitgrid on a fast multicore x64_86 linux server with gobs of RAM.

conda install (recommended)¶

fitgrid is packaged and distributed on Anaconda Cloud for installation into conda virtual environments using the conda package manager. The virtual environment isolates the fitgrid installation to prevent clashes with what is already installed on your system and the conda installer automatically searches for compatible versions of the hundreds of Python and R packages fitgrid requires to run. The steps are simple but before proceeding it is prudent to have a general understanding of conda virtual environments and at least these commands: conda create ..., conda activate ..., conda deactivate, conda install -c ..... See the Conda Cheat Sheet for a summary.

This example shows how to install fitgrid from a bash shell command prompt in a linux terminal window. It assumes the conda executable is already installed (we use miniconda3). It bakes a date into the name and provisions the new environment with fitgrid and a compatible jupyterlab. The order of the -c conda channels matters, the channels are searched for compatible packages from left to right.

Example installation

$ conda create --name fg_012021 -c kutaslab -c defaults -c conda-forge fitgrid jupyterlab
$ conda activate fg_012021

Note

To avoid Python and R package version conflicts, experience teaches it is helpful to start with a fresh conda environment if possible, and install fitgrid along with any other conda packages you want to run in one fell swoop. The conda installer cannot see into the future. If you install packages one by one, the versions you get will be compatible with what you have already installed but may not be compatible with what you want to install next.

pip install¶

This is asking for trouble because Python packaging doesn’t know or care about the R dependencies. The fitgrid source may be downloaded or git cloned from https://github.com/kutaslab/fitgrid and we upload stable releases of the Python package to PyPI (here) as a courtesy, it is not intended for general use.

System requirements¶

The platform of choice is linux. Minimum system requirements are not known but obviously large scale regression modeling with millions of data points is computationally demanding. We develop, test, and use fitgrid on a high-performance linux server (CentOS 7, Intel x86_64, 48 cores, 500GB RAM) and our continuous integration runs all but the parallel processing pytests on Ubuntu 18.04, 7GB RAM. The 64-bit OSX conda package is spot-checked from time to time on an Intel MacBook Pro (8 cores, 32GB RAM). The pytests pass when conda is installed as above, useability for modeling at scale is unknown. We don’t test the 64-bit Windows conda package, field reports from contributors are welcome.