samuela / BCFWstruct

Block-Coordinate Frank-Wolfe Optimization for Structural SVMs

#BCFWstruct

BCFWstruct is a Matlab implementation of the Block-Coordinate Frank-Wolfe solver for Structural SVMs. For more information about the algorithm please check our ICML 2013 paper.

The code is organized as follows:

• solvers contains the optimization methods, including the block-coordinate Frank-Wolfe solver (BCFW). If you want to use BCFW in your project, you most likely only need to run an addpath(genpath('solvers')) in your Matlab sources.
• demos contains the application-dependent code, such as MAP decoding or the feature map computation. The source code includes a sequence prediction demo for optical character recognition (OCR).
• data is initally empty, it is used to store the data files required for the demos.

##Getting Started

1. You need a working installation of Matlab.
2. Clone the git repository: git clone [e-mail]:ppletscher/BCFWstruct.git (or if you don't want to use git, you can download a zip file of the latest version of the code by following this link).
3. Obtain the data files required to run the demos. On Unix systems you can simply run ./fetch_data.sh. On Windows, you can use Cygwin or manually download the listed files and put them in the data folder.
4. For the OCR demo change to demo/chain and run ocr from within Matlab.

##Usage

If you would like to use the BCFW solver for your own structured output prediction problem, you will need to implement three functions:

• The feature map.
• The maximization oracle.
• The loss function.

You can find an example implementation in the demo/chain folder. For an overview of the exact usage and the supported options, please check the Matlab documentation of the solvers.

Note that BCFW uses a similar calling interface as the one from the Matlab wrapper to SVM^struct implemented by Andrea Vedaldi. Users of SVM^struct can thus easily use BCFW with only a tiny change in their code (see the Matlab documentation of solverBCFW for more details).

##Citation

Please use the following BibTeX entry to cite this software in your work:

@inproceedings{LacosteJulien2013,
  author    = {Lacoste-Julien, Simon and Jaggi, Martin and Schmidt, Mark and Pletscher, Patrick},
  title     = {Block-Coordinate {F}rank-{W}olfe Optimization for Structural {SVMs}},
  booktitle = {ICML},
  year      = {2013},
}

##Authors

• Simon Lacoste-Julien
• Martin Jaggi
• Mark Schmidt
• Patrick Pletscher

##Octave Support

The code also works with Octave, this was tested with Octave 3.6.4. In order to get the progress update working, we recommend running more off before calling our structured SVM solvers.