Installation guide

elapid is accessible from pypi:

pip install elapid

It's also accessible from conda:

conda install -c conda-forge elapid

This should suffice for most linux/mac users, as there are builds available for most of the dependencies (numpy, sklearn, glmnet, geopandas, rasterio).

Windows installs can be more difficult. There are two primary challenges you may face: installing the key geospatial dependencies, and installing glmnet.

You can avoid both of them by using Windows Subsystem for Linux, which creates a linux environment on your Windows machine. If you're using an older version of Windows, see the options below.

---

Resolving geospatial dependencies

With conda

If you have conda installed, use that to install elapid's dependencies:

conda create -n elapid python=3.8 -y
conda activate elapid # or just `activate elapid` on windows
conda install -y -c conda-forge geopandas rasterio rtree scikit-learn tqdm
pip install elapid

Installing on Windows without conda

You can get Windows builds of several key geospatial packages using pipwin, which installs wheels from an unofficial source:

pip install wheel pipwin
pipwin install numpy
pipwin install pandas
pipwin install shapely
pipwin install gdal
pipwin install fiona
pipwin install pyproj
pipwin install six
pipwin install rtree
pip install geopandas
pip install elapid

---

Installing glmnet

glmnet needs to be manually installed on Windows. But technically it's not required.

elapid was written to try and match the modeling framework of the R version of Maxent, maxnet. maxnet uses an inhomogeneous Poisson process model, which fits penalized maximum likelihood models, and is handled by the package glmnet.

There is a python wrapper for glmnet, which is used by elapid. But it has no Windows build, so it has to compile some fortran code on install. This means you need to have a fortran compiler running (like MinGW-w64 or Cygwin) if you want to install it (pip install glmnet).

You can also checkout this GitHub issue to read about other people's solutions or contribute a better solution.

An important consideration

To simplify installing and working with elapid, you can install the package and fit Maxent models with or without glmnet. This is handled by fitting the maximum likelihood model with either glmnet or with sklearn.

The results are similar, but not the same. This is largely because of differences in how regularization is handled.

glmnet can handle arrays of regularization terms, which allows fine-scale control over the potential importance of different features. There are a series of calibrated defaults used by elapid, which were originally defined in maxnet but have been updated in opinionated ways.

You can typically just assign a single value for regularization scores with sklearn.

The differences are relatively small. When comparing models fit in maxnet to elapid models, the level of agreement was r2 = 0.91 for glmnet models, r2 = 0.85 for sklearn models.

Still, I recommend users try their best to install glmnet if you're interested in maintaining fidelity to the other family of Maxent tools.