Quick start

Install

REMFlow requires Python 3.11 or newer. The recommended installation uses the released remflow package from PyPI:

python -m pip install remflow

JAX is optional. Install the extra if the analysis will use a JAX backend:

python -m pip install "remflow[gpu]"

The appropriate JAX CUDA build still depends on the operating system, driver, and CUDA installation. Follow the JAX installation instructions before using backend="jax:gpu".

A source checkout is needed only for development. In that case, use:

python -m pip install -e ".[dev,docs]"

Fit a model

Start with a pandas data frame containing an event time, sender, and receiver:

import pandas as pd

from remflow import diagnostics, remify, remstats, remstimate

events = pd.DataFrame(
    {
        "time": [1.0, 2.0, 3.0, 4.0, 5.0],
        "sender": ["A", "B", "A", "C", "B"],
        "receiver": ["B", "A", "C", "A", "C"],
    }
)

history = remify(events, actors=["A", "B", "C"], ordinal=False)
statistics = remstats(
    history,
    tie_effects="~ inertia() + reciprocity() + otp()",
    first=2,
)
fit = remstimate(history, statistics, backend="numpy")

print(fit.summary())
print(diagnostics(fit, history, statistics).summary())

Here, ordinal=False selects an exact-time intensity model. Use ordinal=True for a conditional event-choice model. Public actor and dyad IDs are 1-based, while the original labels remain available on the history object.

Select a JAX device

fit = remstimate(history, statistics, backend="jax:gpu")

This requires a physical JAX GPU device. REMFlow raises BackendUnavailable if none is visible. Use backend="jax:cpu" when the calculation must run with JAX on the CPU. Reported NumPy/JAX comparisons use float64.

Where to go next

The usage guide covers actor models, duration data, sampled risk sets, GPU selection, the general high-level facade, and its misinformation specialization.
The API reference lists public functions and result objects.
Performance evaluation documents the benchmark commands.
The classroom event study explains the tutorial-derived example and its interpretation limits.