API Reference¶

Complete public API surface for torchregress. Every page in this section is a standalone reference — you can read it without any auto-generation tools. For conceptual guides, derivations, and worked examples, see User Guide, Methods & Algorithms, and Examples.

Modules¶

Module	What's in it	Reference
`torchregress.losses`	Regression losses: point, robust, Gaussian, quantile, ordinal, censored, Poisson/Tweedie, EIV, conformal, …	Losses API
`torchregress.models`	Strong-default tabular MLP, robust preprocessing, `fit_tabular` trainer	Models API
`torchregress.estimators`	Experimental model-agnostic `ConformalRegressor`, `CalibratedRegressor` and `calibrated_deep_ensemble`	Estimators API
`torchregress.ensemble`	Deep ensembles, heteroscedastic ensembles, batch ensembles, SWAG, MC-dropout, BNN	Ensemble API
`torchregress.losses.conformal`	Conformal predictors & `ConformalLoss`	Conformal API
`torchregress.algorithms`	IRLS, SIMEX, RC, LatentNN, TicTac, heteroscedastic Laplace, IVON	Algorithms API
`torchregress.metrics`	Point, distribution, interval, calibration, OOD, ensemble, multivariate, ordinal, censored, decision	Metrics API
`torchregress.calibration`	Variance temperature, isotonic mean, PIT, semi-conformal, label shift	Calibration API
`torchregress.constraints`	Output-head constraints (non-negative, bounded, non-crossing, simplex, spectral norm)	Constraints API
`torchregress.inference`	Prediction-powered inference (PPI)	Inference API
`torchregress.causal`	Doubly-robust ATE/CATE, overlap diagnostics, policy value	Causal API
`torchregress.viz`	Diagnostic, monitoring, results, and utility plotting	Visualization API
`torchregress.test_time`	Bayesian linear heads, OT shift conformal, BLR predictive adapters	Test-time API
`torchregress.semi_supervised`	Teacher–student semi-supervised trainer, consensus/agreement, pseudo-loss, trust weighting	Semi-supervised API
`torchregress.comparison`	Reproducible comparison-example helpers (seeds, timing, metrics, JSON summaries, fairness notes)	Comparison API
`torchregress.prediction`	Predictive batch containers (`PredictiveBatch`) and quantile / bar / sample → density-grid conversion	Utilities API
`torchregress.utils`	Tensor ops, validation, augment, ordinal, propensity, transforms	Utilities API
`torchregress.method_catalog`	Scriptable method metadata (`list_methods`, `get_method_metadata`, …)	Utilities API
`torchregress.health`	`check_health()` installation smoke test	Utilities API

Top-level exports¶

Symbol	Description
`losses`	Lazy submodule of regression losses (point, robust, Gaussian, quantile, ordinal, censored, count, EIV, conformal).
`metrics`	Lazy submodule of evaluation metrics (point, distribution, interval, calibration, OOD, ensemble, decision).
`algorithms`	Lazy submodule of error-in-variables and robust algorithms (IRLS, SIMEX, RC, LatentNN, TicTac, IVON).
`models`	Lazy submodule of the tabular MLP, its preprocessor and trainer.
`estimators`	Lazy submodule of experimental model-agnostic conformal / calibration wrappers and the calibrated deep-ensemble recipe.
`ensemble`	Lazy submodule of deep ensembles, MC-dropout, SWAG, snapshot and Bayesian model building blocks.
`semi_supervised`	Lazy submodule of teacher–student semi-supervised training.
`test_time`	Lazy submodule of test-time adaptation (Bayesian heads, label shift, OT conformal reweighting, alignment).
`method_catalog`	Lazy submodule exposing scriptable method metadata.
`inference`	Lazy submodule of prediction-powered inference (PPI) estimators.
`constraints`	Lazy submodule of output-head constraints (non-negative, bounded, non-crossing, simplex, spectral norm).
`calibration`	Lazy submodule of post-hoc calibrators and shift estimators.
`causal`	Lazy submodule of doubly-robust causal estimation and overlap diagnostics.
`prediction`	Lazy submodule of predictive batch containers (`PredictiveBatch`).
`viz`	Lazy submodule of diagnostic, monitoring, results, and utility plotting.
`comparison`	Lazy submodule of reproducible comparison-example helpers (seeds, timing, metrics, JSON summaries).
`utils`	Lazy submodule of tensor ops, validation, augmentations, ordinal helpers, propensity, and transforms.
`__version__`	Installed torchregress version string.

Package structure¶

torchregress/
├── losses/                # Regression and distributional losses
│   ├── base.py            # BaseLoss, RegressionLoss, DistributionLoss, weighted wrappers
│   ├── gaussian.py        # GaussianNLL, FaithfulGaussian, MultivariateGaussian, LowRankGaussian
│   ├── beta_nll.py        # BetaNLL
│   ├── gaussian_wasserstein.py
│   ├── faithful_gaussian.py
│   ├── robust.py          # Huber, Cauchy, Tukey, Barron, Adaptive, CVaR, LogCosh, Charbonnier, PseudoHuber
│   ├── quantile.py        # Quantile, MultiQuantile, QuantileCrossoverLoss
│   ├── expectile.py       # Expectile, MultiExpectile
│   ├── conformal.py       # SplitConformal, CQR, UACQR, Density, MonteCarlo, Local, CTI, SLS, NexCP, MultivariateScore, ConformalLoss
│   ├── censored.py        # CensoredGaussianNLL, CensoredQuantile, AFT
│   ├── poisson.py         # PoissonDeviance, NegativeBinomial, ZeroInflatedPoisson
│   ├── poisson_gaussian.py
│   ├── tweedie.py         # Tweedie, Gamma, InverseGaussian, CompoundPoisson
│   ├── ordinal.py         # CORAL, CumulativeLink, OrdinalCrossEntropy
│   ├── mdn.py              # MixtureDensityLoss
│   ├── nflows.py           # NormalizingFlowLoss, ContrastiveFlowLoss
│   ├── imbalanced.py      # Density, Focal, LDS, Propensity, BMC, BalancedMSE
│   ├── uncertain_gt.py    # NoisyTargetGaussianNLL, PseudoLabel, Consistency
│   ├── eiv.py             # BaseEIVLoss, FunctionalEIV, StructuralEIV, ODR, EnsembleEIV
│   ├── transforms.py      # Log, BoxCox, Sqrt, YeoJohnson, TransformedTarget
│   ├── balanced_mse.py
│   ├── evidential.py      # EvidentialRegression
│   └── loss_registry.py
├── metrics/               # Evaluation metrics
│   ├── point.py           # mse, rmse, mae, r2_score, huber_loss, mape, msle, ev_score
│   ├── distribution.py    # crps_gaussian, gaussian_nll, energy_score, PIT
│   ├── interval.py        # interval_score, PICP, MPIW
│   ├── calibration.py     # ECE, MCE, PIT-based calibration
│   ├── ensemble.py        # GaussianNLLEnsemble, uncertainty_decomposition
│   ├── ood.py             # Mahalanobis, Typicality, Entropy, KDE
│   ├── multivariate.py    # MultivariateRMSE, MultivariateMAE
│   ├── ordinal.py         # ordinal_accuracy, QWK
│   ├── censored.py        # concordance_index, censoring_rate
│   ├── decision.py        # RiskCoverageCurve, RejectionPolicy
│   ├── tac.py             # Task-agnostic correlations
│   └── uncertain.py       # noisy_target_gaussian_nll, consistency_error
├── calibration/           # Post-hoc transforms and shift estimators
├── ensemble/              # Ensemble models and building blocks
├── algorithms/            # IRLS, SIMEX, RC, LatentNN, TicTac, IVON
├── test_time/             # Bayesian linear heads, OT shift, BLR adapters
├── inference/             # Prediction-powered inference
├── constraints/           # Output-head constraints
├── comparison.py          # Comparison-example helpers (seeds, JSON summary)
├── prediction.py          # Predictive batch containers
├── viz/                   # Diagnostic, monitoring, results, utils plots
├── semi_supervised.py     # TeacherStudentTrainer
└── utils/                 # tensor_ops, validation, augment, ordinal, propensity, transforms

Core imports¶

The library uses lazy submodule imports — accessing a submodule via attribute (tr.losses) is the only cost; it does not load the others.

import torchregress as tr

# Top-level, eagerly imported
from torchregress import BaseLoss, DistributionLoss, RegressionLoss, iteratively_reweighted_fit
from torchregress.algorithms import reweight_fixed_predictions

# Lazy-loaded submodules
losses = tr.losses            # tr.losses.WeightedHuberLoss(...)
metrics = tr.metrics          # tr.metrics.mean_squared_error(y_pred, y_true)
ensemble = tr.ensemble        # tr.ensemble.BaseEnsembleModel(...)
calibration = tr.calibration  # tr.calibration.VarianceTemperatureScaler()
viz = tr.viz                  # tr.viz.set_style(); tr.viz.plot_residuals(...)
constraints = tr.constraints
causal = tr.causal
inference = tr.inference
algorithms = tr.algorithms
test_time = tr.test_time
utils = tr.utils
comparison = tr.comparison
prediction = tr.prediction
semi_supervised = tr.semi_supervised
method_catalog = tr.method_catalog  # Scriptable method metadata

Versioning & stability¶

The library is on 0.x — minor versions may include API changes.
Mature/stable APIs are tagged Core in the Method Catalog.
Advanced methods are powerful but require more careful validation before adoption.