API Reference

Complete API documentation for deep-inference.

API Reference

• Inference API
  • API Overview
  • structural_dml() - Legacy API
  • inference() - New Flexible API
  • Configuration Guidelines
  • Algorithm Overview
  • Expected Performance
• Families Module
  • Factory Function
  • Available Families
  • Base Class
  • Family Classes
  • Usage Example
  • Family Methods
• Targets Module
  • Overview
  • Built-in Targets
  • Custom Targets
  • Target Protocol
  • Using Targets with inference()
  • Autodiff Jacobian
• Lambda Strategies
  • Overview
  • Regime Detection
  • Lambda Strategies
  • Treatment Distributions
  • Strategy Selection API
  • Lambda Strategy Protocol
  • Diagnostics
  • When to Use Each Regime
• Models Module
  • Network Classes
  • Network Architecture
  • Usage with structural_dml
  • Training Configuration
  • Architecture Guidelines
  • Custom Network Usage
• Metrics Module
  • Main Functions
  • Usage Example
  • Key Metrics
  • Validation Targets
  • Interpreting Results
  • Monte Carlo Validation

Quick Reference

Two APIs

API	Use Case
structural_dml()	Production, 13 families, fixed target E[β]
inference()	Flexible targets, regime detection, RCT support

Main Entry Points

from deep_inference import structural_dml, inference

# Legacy API (production-ready)
result = structural_dml(
    Y=Y, T=T, X=X,
    family='linear',
    hidden_dims=[64, 32],
    epochs=100,
    n_folds=50
)

# New API (flexible)
from deep_inference.lambda_.compute import Normal

result = inference(
    Y=Y, T=T, X=X,
    model='logit',
    target='ame',              # Flexible target
    is_randomized=True,        # Regime A
    treatment_dist=Normal(0, 1)
)

Available Families

from deep_inference import FAMILY_REGISTRY
print(list(FAMILY_REGISTRY.keys()))
# ['linear', 'logit', 'poisson', 'tobit', 'negbin', 'gamma', 'gumbel', 'weibull',
#  'gaussian', 'probit', 'beta', 'zip', 'multinomial_logit']

Family Classes

from deep_inference import (
    LinearFamily, LogitFamily, PoissonFamily, TobitFamily,
    NegBinFamily, GammaFamily, GumbelFamily, WeibullFamily,
    MultinomialLogitFamily,
)
from deep_inference.families import (
    GaussianFamily, ProbitFamily, BetaFamily, ZIPFamily,
)

Module Overview

structural_dml

The main entry point. Trains a structural neural network with influence function-based inference.

from deep_inference import structural_dml

result = structural_dml(
    Y,                      # Outcome variable (n,)
    T,                      # Treatment variable (n,)
    X,                      # Covariates (n, d)
    family='linear',        # Statistical family
    hidden_dims=[64, 32],   # Network architecture
    epochs=100,             # Training epochs
    n_folds=50,             # Cross-fitting folds
    lr=0.01,               # Learning rate
    batch_size=64,         # Mini-batch size
    weight_decay=1e-4,     # L2 regularization
    verbose=False          # Print progress
)

DMLResult

The result object returned by structural_dml:

Attribute	Description
mu_hat	Debiased point estimate of E[beta(X)]
mu_naive	Naive (biased) estimate
se	Standard error
ci_lower	Lower bound of 95% CI
ci_upper	Upper bound of 95% CI
theta_hat	Estimated parameters (n, theta_dim)
psi	Influence scores (n,)
diagnostics	Dict with training diagnostics

families

Statistical families defining loss functions, gradients, Hessians, and influence scores.

targets

Target functionals for inference: AverageParameter, AME, CustomTarget.

lambda

Lambda estimation strategies: ComputeLambda (Regime A), AnalyticLambda (B), EstimateLambda (C).

models

Neural network architectures: StructuralNet for parameter estimation.

metrics

Helper functions for computing coverage and SE ratios.