Energy-Based Models API Reference

Complete API documentation for energy-based models (EBMs) in Artifex.

Coming Soon

Energy-based model implementations are planned for a future release. This documentation will be updated when the feature is available.

Overview

Energy-based models will include:

• Score Matching: Noise-contrastive estimation
• Contrastive Divergence: CD-k training algorithms
• Stein Variational Gradient Descent: Particle-based inference
• Langevin Dynamics: MCMC-based sampling

Planned API

Base Class

from artifex.generative_models.models.ebm.base import EnergyBasedModel

model = EnergyBasedModel(
    config: EBMConfig,
    *,
    rngs: nnx.Rngs,
)

Configuration

from artifex.generative_models.core.configuration import EBMConfig

config = EBMConfig(
    name="ebm_model",
    energy_net_hidden_dims=(256, 128),
    mcmc_steps=10,
    step_size=0.01,
)

Energy Function

# Define energy function
energy = model.energy(x)  # Lower = more likely

# Compute log probability
log_prob = -energy

Sampling

# Sample using Langevin dynamics
samples = model.sample(
    n_samples=100,
    n_steps=1000,
    step_size=0.01,
)

Related Documentation

• Energy-Based Models Guide - Conceptual overview
• EBM Concepts - Understanding energy-based modeling
• MCMC Sampling - Sampling algorithms for EBMs

References

• LeCun et al., "A Tutorial on Energy-Based Learning" (2006)
• Du & Mordatch, "Implicit Generation and Modeling with Energy-Based Models" (2019)
• Song & Ermon, "Generative Modeling by Estimating Gradients of the Data Distribution" (2019)