TorchDCM

PyTorch-first discrete choice model estimation and econometric inference.

Installation | Quick Start | Examples | Model Zoo | Development

TorchDCM is the importable Python package for discrete choice model estimation. This repository is intentionally package-first: it keeps the reusable torchdcm implementation, unit tests, examples, and packaging metadata in one small repo that users can install and import directly.

The paper's reproducibility repository is separate and contains benchmark runners, aligned datasets, validation utilities, and committed outputs:

torchdcm-evaluation-benchmark: reproduce the validation and benchmark results

Why TorchDCM

Installation

python -m pip install torchdcm

For local development:

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
python -m pip install -e ".[dev]"
pytest

Quick Start

import torch

from torchdcm import Beta, ChoiceDataset, MultinomialLogit, UtilitySpec
from torchdcm.datasets import make_swissmetro_like

df = make_swissmetro_like(n_obs=300, seed=7)
data = ChoiceDataset.from_wide(
    df,
    alternatives=["TRAIN", "SM", "CAR"],
    choice="choice",
    variables={
        "time": {"TRAIN": "time_train", "SM": "time_sm", "CAR": "time_car"},
        "cost": {"TRAIN": "cost_train", "SM": "cost_sm", "CAR": "cost_car"},
    },
    availability={
        "TRAIN": "avail_train",
        "SM": "avail_sm",
        "CAR": "avail_car",
    },
    individual_id="person_id",
)

spec = UtilitySpec()
spec.utility(
    "TRAIN",
    Beta("ASC_TRAIN")
    + Beta("B_TIME", init=-0.01) * "time"
    + Beta("B_COST", init=-0.1) * "cost",
)
spec.utility(
    "SM",
    Beta("B_TIME", init=-0.01) * "time"
    + Beta("B_COST", init=-0.1) * "cost",
)
spec.utility(
    "CAR",
    Beta("ASC_CAR")
    + Beta("B_TIME", init=-0.01) * "time"
    + Beta("B_COST", init=-0.1) * "cost",
)

device = "cuda" if torch.cuda.is_available() else "cpu"
result = MultinomialLogit(spec, device=device).fit(data, cov_type="cluster", groups="person_id")
print(result.summary())

summary() renders an organized console report covering the model and data, convergence diagnostics, fit statistics, inference, alternative shares, and parameter estimates. The same structured report can be inspected as tables or saved as a reproducible artifact directory:

report = result.report(cov_type="cluster", confidence_level=0.95)
parameter_table = report.parameters

result.save_report(
    "outputs/swissmetro_mnl",
    formats=["html", "json", "csv", "latex", "text"],
)

The output directory contains a readable HTML report, a machine-readable JSON record, parameter/covariance/correlation CSV files, a LaTeX fragment, and a plain-text summary. The convergence section records the exact L-BFGS stopping condition, the internal-parameter gradient infinity norm, and that norm divided by the number of observations. Function- or step-tolerance stops are reported as converged when the normalized gradient does not exceed 1e-5; iteration limits, non-finite values, and larger normalized gradients remain warnings.

All estimators accept a standard PyTorch-style device argument. Passing device="cuda" moves estimation, prediction, simulated likelihoods, and covariance calculations for that model to CUDA when your PyTorch installation has GPU support.

Executed Examples

The examples/ directory contains self-contained Jupyter notebooks for every public model family. Each notebook presents the model's mathematical formulation, builds a nontrivial specification, runs full estimation, and retains its rendered HTML report. They were executed on an AMD Ryzen 9 9950X3D CPU (16 cores), 64 GB RAM, and an NVIDIA GeForce RTX 5090 GPU (32 GB VRAM), running Ubuntu 24.04.4 with PyTorch 2.12.1 and CUDA 13.0. The examples automatically select CUDA when it is available and can be changed to CPU by setting device = "cpu".

Model Zoo

The current package includes:

• multinomial logit / conditional logit;
• nested logit and cross-nested logit;
• mixed logit, WTP-space mixed logit, and error-components logit;
• latent-class logit;
• scaled and covariate-scaled multinomial logit;
• ordered logit and ordered probit;
• hybrid choice with latent variables and Gaussian measurement indicators;
• classic, robust, and cluster covariance estimates;
• WTP and elasticity helpers.

Repository Layout

License

TorchDCM is released under the MIT License.

Development

python -m pip install -e ".[dev]"
pytest

Benchmark and validation work should happen in the companion repository:

git clone https://github.com/mbc96325/torchdcm-evaluation-benchmark.git