A Practical Debugging Tool for Training Deep Neural Networks

A better status screen for deep learning.

Installation • Docs • Experiments • License • Citation

---

pip install cockpit-for-pytorch

---

Cockpit is a visual and statistical debugger specifically designed for deep learning. Training a deep neural network is often a pain! Successfully training such a network usually requires either years of intuition or expensive parameter searches involving lots of trial and error. Traditional debuggers provide only limited help: They can find syntactical errors but not training bugs such as ill-chosen learning rates. Cockpit offers a closer, more meaningful look into the training process with multiple well-chosen instruments.

---

Installation

To install Cockpit simply run

pip install cockpit-for-pytorch

Conda environment

For convenience, we also provide a conda environment, which can be installed via the conda yml file. It includes all the necessary requirements to build the docs, execute the tests and run the examples.

Documentation

The documentation provides a full tutorial on how to get started using Cockpit as well as a detailed documentation of its API.

Experiments

To showcase the capabilities of Cockpit we performed several experiments illustrating the usefulness of our debugging tool. The code for the experiments can be found in a separate repository. For a discussion of those experiments please refer to our paper.

License

Distributed under the MIT License. See LICENSE for more information.

Citation

If you use Cockpit, please consider citing:

Frank Schneider, Felix Dangel, Philipp Hennig
Cockpit: A Practical Debugging Tool for Training Deep Neural Networks
arXiv 2102.06604

@misc{schneider2021cockpit,
   title={{Cockpit: A Practical Debugging Tool for Training Deep Neural Networks}},
   author={Frank Schneider and Felix Dangel and Philipp Hennig},
   year={2021},
   eprint={2102.06604},
   archivePrefix={arXiv},
   primaryClass={cs.LG}
}

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog.

Unreleased

1.0.2 - 2021-10-26

Added

• Added references to a separate experiment repository that publishes the code for all experiments shown in the paper.

Fixed

• Protects the batch_grad field in the case where non-SGD is used together with other quantities that free batch_grad for memory performance. [#5, PR]

1.0.1 - 2021-10-13

From this version on, cockpit will be available as cockpit-for-pytorch on PyPI.

Added

• Make library pip-installable as cockpit-for-pytorch [PR]
• Require BackPACK main release [PR]
• Added a savename argument to the CockpitPlotter.plot() function, which lets you define the name, and now the savedir should really only describe the directory. [PR, Fixes #8]
• Added optional savefig_kwargs argument to the CockpitPlotter.plot() function that gets passed to the matplotlib function fig.savefig() to, e.g., specify DPI value or use a different file format (e.g. PDF). [PR, Fixes #10]

Internal

• Fix #6: Don't execute extension hook on modules with children [PR]

1.0.0 - 2021-04-30

Added

• First public release version of Cockpit.

MIT License

Copyright (c) 2019 Frank Schneider, Felix Dangel & Philipp Hennig

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.