Author: Vlad I. Morariu E-mail: morariu(at)umd(.)edu Date: 2007-06-26 Modified: 2014-10-18

FIGTree is a library that provides a C/C++ and MATLAB interface for speeding up the computation of the Gauss Transform. Its authors are Vlad Morariu, Vikas Raykar, and Changjiang Yang. The authors worked under the supervision of Professor Ramani Duraiswami and Professor Larry Davis, at the University of Maryland.

The current version of the library is based on the following paper:

Vlad I. Morariu, Balaji Vasan Srinivasan, Vikas C. Raykar, Ramani Duraiswami, and Larry S. Davis. Automatic online tuning for fast Gaussian summation. Advances in Neural Information Processing Systems (NIPS), 2008.

For more up-to-date information on this library, please go to the FIGTree homepage,

http://www.umiacs.umd.edu/~morariu/figtree/

the GitHub repo for the most recent changes,

https://github.com/vmorariu/figtree

or the SourceForge download page for releases with Windows binaries,

http://sourceforge.net/projects/figtree

This code makes use of the Approximate Nearest Neighbors (ANN) library for the tree-based datastructure, but if the appropriate flags are supplied (see below), the library can compile without ANN. A slightly modified version of ANN (only makefiles are modified) is provided along with the figtree code for convenience. For newer releases of ANN or for any other information regarding ANN, please go to the authors' website:

http://www.cs.umd.edu/~mount/ANN/

Sample code for using the library in Matlab or C/C++ is provided in 'samples' directory.

Documentation for the Matlab interface is available in the *.m stub files in the 'matlab' directory. The Windows MEX files have a *.dll extension since most Matlab versions accept *.dll files (whereas others accept *.mex and not *.mexw32 or vice versa). In case of conflict with the figtree.dll that provides the C interface, the figtree.dll MEX file can be renamed to have either a *.mex or *.mexw32 extension (depending on the version of Matlab).

For a description of available C/C++ functions, see 'include/figtree.h'.

The figtree precompiled binaries (.dll) and lib files are provided in figtree/bin and figtree/lib.

To use the figtree library, you must add {BASE_PATH}/figtree/lib to the lib path, and add figtree.lib as a dependency.

To execute code (not MATLAB) linked against figtree (and ANN), you must make sure that the locations of ann_figtree_version.dll and figtree.dll are in the system path.

NOTE: See below for an example of how the PATH environment variable can be used to add the paths of libraries without having to copy them in the directory of each executable that uses them.

Change the directory to ${FIGTREE_DIR}/python. To build in place, putting the figtree.so (Linux) or figtree.pyd (Windows) in ${FIGTREE_DIR/python, issue the following command:

python setup.py build_ext --inplace

Before running python, make sure ${FIGTREE_DIR}/python is in the PYTHONPATH environment variable. For example if you are using bash, you can use the python extension by starting python as follows (make sure you set the FIGTREE_DIR variable to the figtree root path first):

PYTHONPATH=${FIGTREE_DIR}/python LD_LIBRARY_PATH=${FIGTREE_DIR}/lib python

Another option is to install the python extension in a standard path using the command

python setup.py install

with the "--user" or "--prefix PREFIX_PATH" flags set as needed.

Another option for using figtree in python is the ctypes wrapper provided by Frederik Beaujean:

https://github.com/fredRos/pyfigtree

This has been tested on Matlab 7 with g++ and msvc as a compiler.

1. Set up matlab to use the gcc/msvc compiler by running 'mex -setup'.

2. In Matlab, set the working directory to (FIGTREE_DIR)/matlab/ and run CompileMexFilesUsingMatlab.m.

You can compile the library using the Visual Studio Solution (FIGTREE_DIR)/vs8/figtree.sln. Compiling the entire solution will compile both the C/C++ library (figtree.dll will be in the bin directory, and figtree.lib will be in the lib directory), and the mex files, which will be placed in the matlab directory.

Make sure you first switch to the 'Release' mode, or else figtree will be very slow!

1. Compile FIGTree library as a shared library by issuing command 'make'

   If you instead want to compile the FIGTree library as a static library, then the command is 'make FIGTREE_LIB_TYPE=static'.

2. If shared libraries are created, make sure that they are in the library path. Assuming that FIGTree libraries are in your home directory, one way to do this is to set LD_LIBRARY_PATH:

   setenv LD_LIBRARY_PATH ${HOME}/figtree/lib

   if LD_LIBRARY_PATH was not previously set, or

   setenv LD_LIBRARY_PATH ${LD_LIBRARY_PATH}:${HOME}/figtree/lib

   if LD_LIBRARY_PATH was previously set.

Assume the figtree.dll is located in C:\figtree\bin. To add it to the PATH environment variable (this does not require admin privileges):

1. Right click on 'My Computer'
2. Click on 'Properties'
3. Click on the 'Advanced' tab
4. At the bottom, click on 'Environment Variables'
5. In the upper section 'User variables for (yourusername)' is a list of variables with their values. Variable 'PATH' may or may not be one of them. a) if PATH is not among the user variables (usually TEMP and TMP are present), then you must add it by clicking 'New', entering 'PATH' in the 'Variable name:' field, and then entering '%PATH%; C:\figtree\bin' 'in the Variable value:' field. The '%PATH%' is there so that the figtree directory is APPENDED to the System variables instance of 'PATH' (which you cannot change if you are not an administrator). b) if PATH is among the user variables, select it, and then append your path to the list, using ';' as delimiter.

• TODO: Clean up makefiles for Unix/Linux/Solaris version.
• TODO: estimate avg neighbors using approach of Faloutsos et al (SIGMOD 2000) instead of querying kd-tree on subsampled sources
• TODO: currently, k-center clustering is performed using an N log(K) + K^2 (expected time) approach based on the Gonzales K-center clustering. Use N log(K) approach instead.
• TODO: add function to improve cost prediction by tuning to actual hardware (i.e. floating point operation times vary relative to memory access times on different machines; also exp() takes different number of float ops depending on machine)
• TODO: add memory allocation checks to avoid crashes
• TODO: allow kd-tree, clusters, and coefficients computed on sources to persist between function calls in case user wants to reuse them
• TODO: make some more of the internal parameters easier to modify by users without requring code to be compiled again (such parameters include: amt of subsampling to do for estimating cost, truncation number limit, etc)
• TODO: modify ANN so that it no longer uses global variables (which prevents figtree from being called simultaneously in multiple threads)

• Fixed bug: Added fabs() calls to cluster-wise error bounds to handle negative weights properly.
• Added verbose print statements for method selection.

• Added stdlib.h and string.h includes to for exit(), strcmp(), memset() (they used to be implicitly included by gcc headers, but they are not anymore).
• Modified sample.cpp to show the recommended default way to call figtree, which uses the automatic method and parameter selection approach described in the NIPS 08 paper.

• NEW: Added function that chooses fastest method between Direct, DirectTree, Ifgt, and IfgtTree, making figtree a black-box approach
• Added point-wise and cluster-wise approaches to selecting source/target truncation numbers
• Moved ann-1.1.1 library inside (figtree-dir)/external directory to reduce confusion
• Change mex code to include ANN code (instead of linking to it) to prevent issues with library paths
• Removed many of the mex files that were really just confusing people, and left only figtree, method selection, and k-center clustering mex files as they are most useful.
• Changed interface of the mex files to remove redundant arguments.
• Other improvements (see source files)

• FIX: parameter selection gave bad params since K (number of clusters) was not allowed to equal number of sources

• Initial adaptation from Vikas's initial code
• NEW: integrated Vikas's FIGTree code into library
• NEW: added parameter selection method that does not assume uniformly distributed sources
• NEW: added C/C++ interface, with one function (figtree()) that does most work
• Sources and targets no longer need to fit in unit hypercube