Skip to main content

ThermodynamicAnalyticsToolkit - analyze loss manifolds of neural networks

Project description

Thermodynamic Analytics Toolkit is a sampling-based approach to understand the effectiveness of neural networks training and investigate their loss manifolds.

It uses Tensorflow (https://www.tensorflow.org/) as neural network framework and implements advanced sampling algorithms on top of it. It contains both a rapid prototyping platform for new sampling methods and also an analysis framework to understand the intricacies of the loss manifold in terms of averages, covariance, diffusion maps, and free energy.

Please take a look at the extensive [userguide](https://alan-turing-institute.github.io/ThermodynamicAnalyticsToolkit/).

Dependencies

In total, we depend on the following python packages:

  • tensorflow (1.4.1, 1.6-1.10; 1.5 is not recommended)
  • numpy
  • pandas
  • scipy
  • scikit-learn
  • acor (see the userguide for installation instructions)

Furthermore, for installation from a cloned git repository or a pure source tarball, the following non-python packages are required for creating all userguides,

  • doxygen,
  • asciidoc, dblatex
  • pdflatex,

and for running all tests,

  • awk, sqlite3.

Finally, there are some optional python packages:

  • pydiffmap: allows diffusion map analysis through pydiffmap package
  • tqdm: allows displaying progress bar during training and sampling

Installation

Use one of the following ways:

For more information please refer to the userguide (see the [releases](https://github.com/alan-turing-institute/ThermodynamicAnalyticsToolkit/releases) on github or [as html version](https://alan-turing-institute.github.io/ThermodynamicAnalyticsToolkit/)) for installation instructions.

Alternatively, the userguide PDF is also contained in the release tarballs in folder doc/userguide. As a fall-back the asciidoc userguide files reside in doc/userguide and are perfectly human-readable, see doc/userguide/introduction.txt. As a last fall-back have a look at INSTALL for general instructions on how to installing a package maintained by autotools, automake.

When cloning from github please call the ./bootstrap.sh script (requiring installed autotools and automake packages).

NOTE: If you only want to use the package and do not plan to submit code, it is strongly advised to use the PyPI package (using `pip`) or the “release” tarballs instead of cloning the repository directly.

Documentation

In general, the documentation is maintained in the folder doc. The asciidoc userguide files reside in doc/userguide and are human-readable in your preferred editor if every other option fails.

There are multiple guides to help you:

  • Userguide: user manual on how to install and use TATi
  • Programmer’s guide: manual on basic programming with Tensorflow and TATi
  • API reference: doxygen-generated API reference

After installation (configure, make, make doc, make install) these guides can be found in the typical documentation directory (e.g., share/doc/thermodynamicanalyticstoolkit/ depending on your OS).

Note that all of the above guides are also available as html versions after installation.

Acknowledgments

TATi has received financial support from a seed funding grant and through a Rutherford fellowship from the Alan Turing Institute in London (R-SIS-003, R-RUT-001), from an EPSRC grant no. EP/P006175/1 (Data Driven Coarse Graining using Space-Time Diffusion Maps, B. Leimkuhler PI), and also from a Microsoft Azure Sponsorship (MS-AZR-0143P).

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Filename, size & hash SHA256 hash help File type Python version Upload date
tati-0.9.5-py3-none-any.whl (235.8 kB) Copy SHA256 hash SHA256 Wheel py3

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN SignalFx SignalFx Supporter DigiCert DigiCert EV certificate StatusPage StatusPage Status page