7. Available Models and Targets

The deep-inference package provides a growing library of built-in model families (see Setup) and target functionals (see Target Functionals). Together, the choice of model (the loss \(\ell\)) and target (the functional \(H\)) fully specify an inference problem.

Custom losses

For models not in the built-in library, users can supply a custom loss function. The package then determines automatically whether the Hessian depends on \(\theta\) (triggering Regime C and three-way splitting) and computes all required derivatives — score, Hessian, and target Jacobian — via automatic differentiation. The user needs only the loss function and the target.

import torch
from deep_inference import inference

def my_loss(y, t, theta):
    """Custom structural loss (Poisson with log-link)."""
    mu = torch.exp(theta[0] + theta[1] * t)  # log-link
    return mu - y * torch.log(mu)            # Poisson NLL

result = inference(
    Y, T, X,
    loss=my_loss, theta_dim=2,
    target_fn=lambda x, th, tt: th[1],
)

Putting it together

The three earlier phases combine into one recipe:

Pick a model — the structural loss \(\ell\) (Setup).
Pick a target — the functional \(H\) (Target Functionals).
The package fits \(\hat\theta(X)\), applies the influence function correction, selects the Lambda regime automatically, and returns \(\hat\mu\) with a valid standard error and confidence interval.