nshdesai / torch-uncertainty Goto Github PK

Torch Uncertainty is a package designed to help you leverage uncertainty quantification techniques and make your neural networks more reliable. It is based on PyTorch Lightning to handle multi-GPU training and inference and automatic logging through tensorboard.

This package provides a multi-level API, including:

• ready-to-train baselines on research datasets, such as CIFAR and ImageNet
• baselines available for training on your datasets
• layers available for use in your networks

This package also contains the official implementation of Packed-Ensembles.

If you find the corresponding models interesting, please consider citing our paper:

@inproceedings{laurent2023packed,
    title={Packed-Ensembles for Efficient Uncertainty Estimation},
    author={Laurent, Olivier and Lafage, Adrien and Tartaglione, Enzo and Daniel, Geoffrey and Martinez, Jean-Marc and Bursuc, Andrei and Franchi, Gianni},
    booktitle={ICLR},
    year={2023}
}

The package can be installed from PyPI or from source.

Install the package via pip:

pip install torch-uncertainty

Installation guidelines for poetry are available on https://python-poetry.org/docs/. They boil down to executing the following command:

curl -sSL https://install.python-poetry.org | python3 -

Clone the repository:

git clone https://github.com/ENSTA-U2IS/torch-uncertainty.git

Create a new conda environment and activate it with:

conda create -n uncertainty && conda activate uncertainty

Install the package using poetry:

poetry install torch-uncertainty

or, for development,

poetry install torch-uncertainty --with dev

Depending on your system, you may encounter errors. If so, kill the process and add PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring at the beginning of every poetry install commands.

In case that you would like to contribute, install from source and add the pre-commit hooks with pre-commit install

Please find the documentation at torch-uncertainty.github.io.

A quickstart is available at torch-uncertainty.github.io/quickstart.

To date, the following baselines are implemented:

• Deep Ensembles
• Masksembles
• Packed-Ensembles

You may find a lot of information about modern uncertainty estimation techniques on the Awesome Uncertainty in Deep Learning.