Installation and reproducibility

Install the research library

Trex is distributed from apoorvalal/trex. Do not run pip install trex: that PyPI name is an unrelated project. Python 3.10 or later is required.

Choose an environment with the appropriate PyTorch build for your CPU or accelerator first. Then install this repository:

git clone https://github.com/apoorvalal/trex.git
cd trex
python -m pip install -e .

The editable installation is convenient when modifying methods or running the included notebooks. A non-editable installation needs no permanent checkout:

python -m pip install "git+https://github.com/apoorvalal/trex.git"

For a reproducible experiment, replace the moving branch with a full commit hash. For example, the correctness-update snapshot is:

python -m pip install "git+https://github.com/apoorvalal/trex.git@98df3cb7880df8c2b61ed0594c43c96e16bee93f"

This pins Trex, not its dependencies. Record the Python, Torch, CUDA/driver (if used), NumPy and SciPy versions too. Current version metadata remains 0.1.0, so a Git revision identifies development snapshots more precisely than that version string.

Runtime and optional dependencies

The core installation includes Torch, NumPy, SciPy, pandas and pytorch-minimize. Many methods use Torch throughout; GEL and score estimation also use CPU-side numerical routines. Installing Trex does not turn every routine into a GPU kernel.

# Independent estimator references and tests
python -m pip install -e '.[test]'

# Quarto execution and API documentation tools
python -m pip install -e '.[docs]'

The Quarto CLI is a separate installation. Large-model generation uses optional Transformers/PEFT/quantization tooling described in its guide; those dependencies and model weights are not required for regression, choice or panel estimation.

Verify the environment

Show Python

import platform
import torch
import trex

print({"python": platform.python_version(), "torch": torch.__version__,
       "trex": trex.__version__, "cuda_available": torch.cuda.is_available()})

{'python': '3.12.14', 'torch': '2.14.1+cpu', 'trex': '0.1.0', 'cuda_available': False}

The executable guides use device="cpu" deliberately, so they do not require a GPU. This does not imply CPU and CUDA are equally fast or bitwise identical. See numerical conventions before moving a workload between devices.