Implementations of quasi-hyperbolic optimization algorithms.
Clone or download
Pull request Compare This branch is 5 commits behind facebookresearch:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
docs
docs_src
qhoptim
test_qhoptim
.flake8
.gitignore
CODE_OF_CONDUCT.rst
CONTRIBUTING.rst
LICENSE
README.rst
pyproject.toml
setup.py

README.rst

qhoptim: Quasi-hyperbolic optimization

This repository contains PyTorch and TensorFlow implementations of the quasi-hyperbolic momentum (QHM) and quasi-hyperbolic Adam (QHAdam) optimization algorithms from Facebook AI Research.

Documentation

Please refer to the documentation for installation instructions, usage information, and an API reference.

Direct link to installation instructions: here.

Reference

QHM and QHAdam were proposed in the ICLR 2019 paper "Quasi-hyperbolic momentum and Adam for deep learning". We recommend reading the paper for both theoretical insights into and empirical analyses of the algorithms.

If you find the algorithms useful in your research, we ask that you cite the paper as follows:

@inproceedings{ma2019qh,
  title={Quasi-hyperbolic momentum and Adam for deep learning},
  author={Jerry Ma and Denis Yarats},
  booktitle={International Conference on Learning Representations},
  year={2019}
}

Contributing

Bugfixes and contributions are very much appreciated! Please see CONTRIBUTING.rst for more information.

License

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.