falcon-sbi 0.4.2

CLI-driven framework for simulation-based inference with large simulators

Falcon

Falcon (Factorized Adaptive Learning of Conditional Orchestrated Networks) is a CLI-driven Python framework for simulation-based inference (SBI) with large, expensive simulators. Born in astrophysics, built for any domain with complex forward models — break your model into components and Falcon jointly infers their parameters.

• Composable — define multi-component models as a graph of simulators in YAML, each wrapped with a thin Python interface, regardless of framework.
• Adaptive — steers simulations toward high-posterior regions as training progresses, focusing compute where it matters.
• Concurrent — trains neural posterior estimators across heterogeneous parameter blocks in parallel, using Ray for distributed execution.
• Batteries included — ships with neural spline flows, data embeddings (including CNN/transformer support), and built-in experiment tracking via WandB.

Installation

pip install falcon-sbi

For development:

git clone https://github.com/cweniger/falcon.git
cd falcon
pip install -e ".[monitor]"

Quick Start

Run the minimal example (a 3-parameter Gaussian inference problem):

cd examples/01_minimal
falcon launch -o output/run_01
falcon sample posterior -o output/run_01

This trains a neural posterior estimator on simulated data, then draws 1000 posterior samples. Results are saved under output/run_01/.

How It Works

You define a directed graph of random variables in config.yml. Each node has a simulator (forward model) and optionally an estimator (learned posterior). Falcon iterates between simulating data and training the estimator, automatically managing the sample buffer.

graph:
  z:                                    # Latent parameters
    evidence: [x]
    simulator:
      _target_: falcon.priors.Hypercube
      priors:
        - ['uniform', -5.0, 5.0]
    estimator:
      _target_: falcon.estimators.Flow

  x:                                    # Observations
    parents: [z]
    simulator:
      _target_: model.Simulate
    observed: "./data/obs.npz['x']"

CLI

falcon launch [-o DIR] [-c CONFIG] [key=value ...]
falcon sample prior|posterior|proposal -o DIR
falcon graph                            # Visualize graph structure
falcon monitor                          # Real-time TUI dashboard (requires pip install "falcon[monitor]")

Examples

Example	Description
01_minimal	Basic 3-parameter inference
02_bimodal	10D bimodal posterior with training strategies
03_composite	Multi-node graph with image embeddings
04_gaussian	Gaussian inference
05_linear_regression	Linear regression

Documentation

For tutorials, configuration reference, and API docs, see cweniger.github.io/falcon.

Citation

@software{falcon2024,
  title = {Falcon: Distributed Dynamic Simulation-Based Inference},
  author = {Weniger, Christoph},
  year = {2024},
  url = {https://github.com/cweniger/falcon}
}

License

MIT License - see LICENSE for details.