Skip to content

Commit

Permalink
Public release announcement.
Browse files Browse the repository at this point in the history
  • Loading branch information
rdadolf committed Sep 27, 2016
1 parent 8617564 commit 62f6dbc
Showing 1 changed file with 23 additions and 4 deletions.
27 changes: 23 additions & 4 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,14 +2,18 @@

[![build status](https://travis-ci.org/rdadolf/fathom.svg?branch=master)](https://travis-ci.org/rdadolf/fathom)

## Release: `0.9-soft`

This release reflects the state of Fathom more or less as it was for the paper published paper in September 2016. We are currently developing a somewhat more user-friendly version, which you can track in the GitHub issue tracker. If you're eager to use Fathom as it is, please let us know. We might be able to help you get started.

## Workloads

This paper contains a full description of the workloads, performance characteristics, and the rationale behind the project:

> R. Adolf, S. Rama, B. Reagen, G.Y. Wei, D. Brooks. "Fathom: Reference Workloads for Modern Deep Learning Methods."
[(Arxiv)](http://arxiv.org/pdf/FIXME.pdf)
[(DOI)](http://dx.doi.org/10.1109/IISWC.2016.FIXME)

## Workloads

Name | Description
-------- | -----
Seq2Seq | Direct language-to-language sentence translation. State-of-the-art accuracy with a simple, language-agnostic architecture.
Expand All @@ -21,11 +25,26 @@ VGG | Image classifier demonstrating the power of small convolutional filte
AlexNet | Image classifier. Watershed for deep learning by beating hand-tuned image systems at ILSVRC 2012.
DeepQ | Atari-playing neural network from DeepMind. Achieves superhuman performance on majority of Atari2600 games, without any preconceptions.

## Data

Fathom does not come with datasets suitable for training. This is a combination of size (realistic training sets are often massive) and licensing (an oft-repeated mantra is that good data is more valuable than a good model).
Regardless, the inputs Fathom is designed for are standard and widely-available.

These links should take you to the original data owners:

- [ImageNet](http://www.image-net.org/download-images) - requires registration, but downloads are free for non-commercial purposes.
- [WMT15](http://www.statmt.org/europarl/)
- [bAbI](https://research.facebook.com/research/babi/)
- MNIST - automatically downloaded by Fathom.
- TIMIT - requires membership of the Linguistic Data Consortium (this is not free, but it is widely available in the research community).
- Atari "Breakout" ROM - available online

## Running

#### Prerequisites:
- TensorFlow v0.8+
- ...
Fathom is tested with TensorFlow v0.8rc0, and due to API instability, there are issues with recent versions (Google has changed TF's layout several times). If you're willing to rename your functions and swap a couple import statements, recent versions of TF should work.

Many of the models require external python libraries (e.g., [ALE](https://github.com/mgbellemare/Arcade-Learning-Environment) for DeepQ). Most of these are available as `pip` packages.

#### Running a single model:

Expand Down

0 comments on commit 62f6dbc

Please sign in to comment.