Multitask Representation Learning

This code was used to produce results reported in the following paper:

Maksim Lapin, Bernt Schiele and Matthias Hein
Scalable Multitask Representation Learning for Scene Classification
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014

The software was tested on Debian GNU/Linux 7.4 (wheezy) using MATLAB R2013a and GCC 4.4.

Getting started

git clone https://github.com/mlapin/cvpr14mtl.git

At MATLAB prompt:

showresults

Playing with the precomputed kernels

Download the precomputed kernels:

cd matlab && make playkernels

At MATLAB prompt:

playground

You may need to recompile the STL-SDCA and MTL-SDCA solvers, see below for instructions.

To download more kernels (excluding the ones from Xiao et al.), run:

cd matlab && make allkernels

Running experiments

STL-SDCA and MTL-SDCA solvers (mex code)

cd mtlsdca && make clean && make

If MATLAB is not found, edit the Makefile and set the path manually,
e.g. MATLAB_PATH = /usr/lib/matlab-8.1

If Intel MKL is installed, specify the corresponding path in INTEL_MKL_PATH;
otherwise, MATLAB BLAS will be used.

To disable verbose output from the solvers, comment out the following line in the Makefile and recompile: STD_CXXFLAGS += -DVERBOSE

USPS/MNIST experiments

cd usps && make

This will compile MATLAB code experiments.m and create a text file cmd_experiments.txt with commands that can be executed in parallel. MCR environment needs to be set up to run the commands, see run_experiments.sh for details. To learn more about working with the compiled MATLAB code, visit
http://www.mathworks.com/help/compiler/working-with-the-mcr.html

SUN397 experiments

First, create the 10 splits. Go to matlab/splits and run at MATLAB prompt:

splits

Next, have a look at the Makefile:

cd matlab && make

This will show a list of available make targets. Note: you must modify the Makefile. At the very minimum, you must specify:

SUN397 = the path to the downloaded SUN397 dataset;
SUN397R100K = the path to a directory where the processed (resized) images will be stored;
arguments to the make/make_cmd_[mtl]experiments.sh scripts, see ex and exmtl make targets.

To resize images to at most 100K pixels, run

make r100k

To run single task learning (STL) experiments, use

make ex

This will compile MATLAB code and create a number of text files with commands that can be executed in parallel. As with the USPS/MNIST experiments, MCR environment needs to be set up to run the commands.

Similarly, to run multitask learning (MTL) experiments, use

make exmtl

Note: all results (precomputed kernel matrices, trained models, test scores, etc.) will be stored in matlab/experiments and will require disk space on the order of 500-700GB. By default, caching of image descriptors (Fisher Vector) is disabled (doNotCacheDescriptors = true in matlab/recognition/traintest.m) and only the kernel matrices are saved to disk. Otherwise, the disk space requirements increase to up to 3-4TB.

mlapin / cvpr14mtl