leap

leap: meta-gradient path learner in MXNet

---

Introduction

This is the MXNet implementation of "leap", the meta-gradient path learner published in ICLR 2019: (link) by S. Flennerhag, P. G. Moreno, N. Lawrence, A. Damianou.

Please cite our work if you find it useful:

@inproceedings{
   flennerhag2018transferring,
   title={Transferring Knowledge across Learning Processes},
   author={Sebastian Flennerhag and Pablo Garcia Moreno and Neil Lawrence and Andreas Damianou},
   booktitle={International Conference on Learning Representations},
   year={2019},
   url={https://openreview.net/forum?id=HygBZnRctX}
}

Getting Started

Check the demos folder. The .ipynb file can be run in Jupyter lab or notebook.

Installation

• Dependencies: Primary dependency is MXNet >=1.2. See all requirements in setup.py.
• Supported architectures / versions: Python 3.6+ on MacOS and Amazon Linux.

• Install from source: To install leap from source, after cloning the repository run the following from the xfer/leap directory:

pip install .

Alternatively, you can install with:

python setup.py install

For installing in editable/development mode, the above commands become respectively: pip install -e . and python setup.py develop .

To confirm installation, run:

>>> import leap
>>> leap.__version__

And confirm that version returned matches the expected package version number.

License

leap is licensed under the Apache 2.0 License.