ppletscher/lpqp

LPQP FOR MAP INFERENCE
----------------------

This is the code implementing the LPQP algorithms introduced in:

Patrick Pletscher & Sharon Wulff
LPQP for MAP: Putting LP Solvers to Better Use
ICML, 2012.

The code was slightly modified after ICML and hence the results might differ a
bit. We re-tested the grid experiments with this version of the code, and
the results pretty much match the ones in the ICML paper. We did not
re-run the protein & DTF experiments, but would expect roughly the same
results as in the published paper.

The tree-based LPQP weighting is implemented in LPQPSDD, the uniform weighting
in LPQPNPBP.


COMPILATION & INSTALL
---------------------

The software was successfully compiled and tested on Ubuntu 12.04, CentOS 6
and Mac OS X 10.8.

Compilation requires cmake, please make sure you have cmake installed on your
system.

1. create a directory build/
2. run ./fetch_external.sh
3. cd to build
4. run cmake
5. run make install
6. run make test to check whether everything runs as expected

the library and the matlab wrappers are now installed into bin/

To use the wrapper in your own scripts, add the directory containing the
mex_lpqp.mex* to the path within Matlab (using addpath).

Remark: Make sure that the compiler optimizations are turned on (-O3),
otherwise the code is *very* slow (probably has to do with the extensive use
of Eigen).


CITATION
--------

If you find the software useful, then please cite the following publication
in your own work:

@inproceedings{Pletscher2012,
  author    = {Pletscher, Patrick and Wulff, Sharon},
  title     = {LPQP for MAP: Putting LP Solvers to Better Use},
  booktitle = {ICML},
  year      = {2012},
}


COPYRIGHT
---------

The LPQP, dual decomposition and tree inference code is written & copyrighted
by Patrick Pletscher and Sharon Wulff.

The TRWS code is written by Vladimir Kolmogorov, and is here just
re-distributed to easily compare to the results obtained using it. See its
folder for the license.