Get startedInstall and extras

Install and extras

Install ergodic from PyPI, add the extras your work needs, and check what your environment has. The core install needs only numpy, pandas and networkx.

Install the core#

pip install "ergodic"

uv add "ergodic"

The core covers graphs, domain knowledge, identification, discovery, effect estimation, process intelligence and the bundled datasets. It never imports an optional package and makes no network calls.

Add extras#

Each extra adds the packages one part of ergodic needs. Quote the name, so your shell does not read the brackets as a pattern:

pip install "ergodic[forecast]"

uv add "ergodic[forecast]"

bayes, scm, forecast and latent add the same package; any one of them is enough for all four. Combine extras with commas: pip install "ergodic[scm,welt]".

Check your environment#

Run this program to print the installed version and the optional packages it finds. The recording shows only the first line, because the rest depends on what you installed.

import importlib.util

import ergodic

# 1. Print the installed version and the graph format it writes.
print("ergodic", ergodic.__version__, "writes", ergodic.GRAPH_FORMAT)

# 2. Check which optional packages this environment has: found or missing.
optional = {
    "pymc": "bayes, scm, forecast and latent",
    "jax": "faster sampling, with numpyro",
    "numpyro": "faster sampling, with jax",
    "welt": "welt",
}
for module, use in optional.items():
    status = "found" if importlib.util.find_spec(module) else "missing"
    print(f"{module:8} {status:8} {use}")

ergodic 0.2.1 writes ergodic.graph/1

• Step 1. The version, and ergodic.graph/1: the format MixedGraph.to_dict() writes. Welt reads this format to convert its graphs.
• Step 2. One line per optional package, such as welt missing welt. missing means you have not installed the extra in the last column.

Speed up sampling#

Forecasting and hidden-regime models sample with PyMC. When jax and numpyro are installed, they sample through PyMC's JAX backend, which ergodic's own measurement found about four times faster on the forecasting model. Without them, sampling falls back to PyMC's C backend and warns once. Neither extra installs them:

pip install jax numpyro

Use your own learners#

Effect estimation takes any model with scikit-learn's fit and predict methods as a learner, and a classifier adds predict_proba. ergodic ships its own OLS and logistic regression and never imports scikit-learn. Install scikit-learn or xgboost yourself to use their models (the welt extra brings scikit-learn in).

Next steps#