Session initialization

Every BI session must start by importing bf and initializing the BI object. This single object is the gateway to everything: data handling, model definition, fitting, diagnostics, and saving.

Rule: Always import bf before importing jax or any other JAX-dependent module, so that the XLA configuration is applied first.

Python

from BayesForge import bf

m = bf(
    platform="cpu",          # "cpu", "gpu", or "tpu"
    cores=None,              # CPU cores (None = all available)
    float_precision=64,      # 64 for float64, 32 for float32
    rand_seed=42,            # int for reproducibility, True for random
    gpu_index=None,          # GPU index when platform="gpu"
    deallocate=False,        # deallocate existing device first
    print_devices_found=True,# print device info on init
)

Arguments

Argument	Type	Default	Description
`platform`	str	`"cpu"`	Hardware backend: `"cpu"`, `"gpu"`, or `"tpu"`
`cores`	int	`None`	CPU cores to allocate (`None` = all). Only for `platform="cpu"`
`float_precision`	int or str	`64`	JAX floating-point precision. Pass `64` (`"float64"`) or `32` (`"float32"`)
`rand_seed`	int or bool	`True`	Reproducibility seed. Pass an `int` (e.g. `42`) for reproducible results, or `True` for entropy-based randomness
`gpu_index`	int	`None`	Which GPU to use by index. Only for `platform="gpu"`
`deallocate`	bool	`False`	Deallocate existing device before setting up new configuration
`print_devices_found`	bool	`True`	Print detected devices on initialization
`backend`	str	`"numpyro"`	Inference backend: `"numpyro"` or `"tfp"`

Common initialization patterns

# Reproducible CPU session (default for development)
m = bf(platform="cpu", rand_seed=42, float_precision=64)

# GPU with specific device
m = bf(platform="gpu", gpu_index=0, rand_seed=42)

# Multi-core CPU for parallel chains
m = bf(platform="cpu", cores=8, rand_seed=42)

# 32-bit precision for memory efficiency
m = bf(platform="cpu", rand_seed=42, float_precision=32)

R

library(BayesianInference)
m = importBI(platform = "cpu", rand_seed = 42)

Julia

using BayesianInference
m = importBI(platform = "cpu", rand_seed = 42)

Saving and reusing a BI object

After fitting a model, save the complete BI object to disk. This persists the model, data, posteriors, and full sampler state — allowing you to reload and analyze later without re-fitting.

Save

m.save("/path/to/my_model.pkl")
# Default filename: "{model_name}_bf.pkl" in current directory

Load

m = bf.load("/path/to/my_model.pkl")

Once loaded, you can access:

m.posteriors        # Posterior samples
m.posteriors_full   # Full posterior (before any filtering)
m.summary()         # Posterior summary table
m.data_on_model     # Data used in the fit
m.model             # Model function
m.diag              # Diagnostic tools

Workflow with save/load

# --- Session 1: Fit and save ---
from BayesForge import bf
m = bf(platform="cpu", rand_seed=42)

def model(x, y):
    alpha = m.dist.normal(0, 10, name="alpha")
    beta  = m.dist.normal(0, 1, name="beta")
    sigma = m.dist.exponential(1, name="sigma")
    m.dist.normal(alpha + beta * x, sigma, obs=y)

m.fit(model, obs=dict(x=x, y=y))
m.save("/tmp/linear_regression.pkl")

# --- Session 2: Load and analyze (no re-fit) ---
from BayesForge import bf
m = bf.load("/tmp/linear_regression.pkl")
print(m.summary())
print(m.posteriors["beta"].mean())

Note: The object is serialized with cloudpickle, so the full model closure and sampler state are preserved. The saved .pkl file is self-contained and portable.