sbi

The new training infrastructure is backward compatible. You can continue using the old method of passing individual keyword arguments to .train(), or you can mix the old style with new features like LoggingConfig or EarlyStopping during a transition period.

# Old API still supported - no breaking changes
inference.train(training_batch_size=100, learning_rate=1e-3)

# Mix old and new as needed during migration
inference.train(
    training_batch_size=100,  # old style
    logging=LoggingConfig(backend="wandb")  # new feature
)

sbi supports MCMC samplers from Pyro and PyMC. These are optional and must be installed explicitly using extras or separate package names depending on your package manager.

Using uv:

• For Pyro samplers (HMC, NUTS): uv pip install "sbi[pyro]"
• For PyMC samplers (HMC, NUTS, Slice): uv pip install "sbi[pymc]"
• For all optional dependencies: uv pip install "sbi[all]"

Using pip:

• For Pyro samplers: python -m pip install "sbi[pyro]"
• For PyMC samplers: python -m pip install "sbi[pymc]"
• For all optional dependencies: python -m pip install "sbi[all]"

Using conda:

• For Pyro samplers: conda install --channel conda-forge pyro-ppl
• For PyMC samplers: conda install --channel conda-forge pymc

The recommended way to install sbi is using uv. sbi requires Python 3.10 or higher, though Python 3.12 is recommended.

To set up a new environment and install sbi with uv:

1. Create a virtual environment with Python 3.12: uv venv --python 3.12
2. Activate the environment:
   • macOS/Linux: source .venv/bin/activate
   • Windows: .venv\Scripts\activate
3. Install the package: uv pip install sbi

# Create a virtual environment with Python 3.12
uv venv --python 3.12

# Activate the environment (on macOS/Linux)
source .venv/bin/activate

# Install sbi
uv pip install sbi

To install sbi directly using pip, use the following command. Ensure you are using Python 3.10 or higher.

python -m pip install sbi

To contribute to sbi, follow these steps to set up your local environment using uv and git:

1. Fork the repository on GitHub.
2. Clone your fork locally:

   git clone git@github.com:$USERNAME/sbi.git
   cd sbi

3. Install in editable mode with development dependencies using uv. This ensures changes to the source code are immediately reflected when importing sbi:

   uv pip install -e ".[dev]"

4. Add the upstream remote to keep your fork synchronized with the main repository:

   git remote add upstream git@github.com:sbi-dev/sbi.git

5. Install pre-commit to run code style checks automatically:

   pre-commit install

You are permitted to use AI coding assistants (e.g., GitHub Copilot, ChatGPT) when contributing to sbi, provided you follow these responsibilities:

• Verification: Do not blindly accept suggestions. You are responsible for the correctness, efficiency, and security of all submitted code.
• Review: Treat AI-generated code as unverified code; review it line-by-line.
• Testing: Write comprehensive unit tests to guard against AI hallucinations or subtle bugs.
• Attribution: If an AI assistant generated a significant portion of the code, mention it in your commit message or pull request description (e.g., "Implemented feature X with assistance from GitHub Copilot").
• Style: Ensure all AI-generated code adheres to the project's formatting and maintainability standards.

To develop documentation with automatic rebuilding and live serving, use sphinx-autobuild. This method is recommended because it automatically runs Jupyter notebooks, collects their output, and incorporates them into the documentation. The server typically runs at http://127.0.0.1:8000.

# Install documentation dependencies
uv sync --extra doc

# Start development server with autoreload
sphinx-autobuild . _build/html

Install the core sbi package using uv. Note that sbi requires Python 3.10 or higher.

To install optional MCMC samplers from Pyro or PyMC, use the following extras:

• sbi[pyro]: For Pyro samplers (e.g., HMC, NUTS).
• sbi[pymc]: For PyMC samplers (e.g., HMC, NUTS, Slice).
• sbi[all]: Installs both Pyro and PyMC support.

uv pip install sbi

# Optional samplers
uv pip install "sbi[pyro]"
uv pip install "sbi[pymc]"
uv pip install "sbi[all]"

To add sbi to a project managed by pixi, use the pixi add command.

pixi add sbi

To ensure sbi is installed correctly, run the minimal example provided in the package. This should return a posterior object without errors.

from sbi.examples.minimal import simple
posterior = simple()
print(posterior)