shuai-huang / dr-mim Goto Github PK

Author: Shuai Huang, The Johns Hopkins University.
Email: [email protected]
Homepage: https://sites.google.com/site/shuangsite/

Last change: 09/22/2016
Change log: 
    v1.0 (SH) - First release (09/22/2016)
    

----------------------------------------------------------------
This package contains source code for performing the supervised dimensionality reduction approach described in the following paper:

@INPROCEEDINGS{MIM2016,
author={S. Huang and T. D. Tran},
booktitle={2016 IEEE International Conference on Image Processing (ICIP)},
title={Dimensionality reduction for image classification via mutual information maximization},
year={2016},
pages={509-513},
month={Sept},
}

If you use this code and find it helpful, please cite the above paper. Thanks:)
----------------------------------------------------------------

The code is written in C++, and requires the following additional libraries:
    a) Eigen (C++ library) http://eigen.tuxfamily.org/
    b) Spectra: C++ Library For Large Scale Eigenvalue Problems https://github.com/yixuan/spectra/
    c) OpenBLAS: An optimized BLAS library http://www.openblas.net/
    
For convenience, Eigen, Spectra, OpenBLAS are all included in this package.
-----------------------------------------------------------------

****************
* Installation *
****************
Please follow the following steps:

0. Go to the directory where the readme file is.

1. Eigen and Spectra do not need additional installation steps. OpenBLAS should be complied and installed in the UNIX environment.
    a) Go to OpenBLAS directory: cd ./OpenBLAS-0.2.19
    b) Compile: make
    c) Install: sudo make install 
       
2. Compile the source code for dimensionality reduction. By default, the OpenBLAS is installed to the directory /opt/OpenBLAS . Should you choose your own directory, please use the correct path accordingly.
    
    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/OpenBLAS/lib/
    g++ -fopenmp -O3 -o main main.cpp DataReader.cpp DimRed.cpp DFE.cpp global.cpp -I ./eigen-eigen-3.3.3 -I ./spectra-0.5.0/include -I /opt/OpenBLAS/include/ -L /opt/OpenBLAS/lib/ -lopenblas
       
**********************
* Running Experiment *
**********************

The experiments can be run using the bash script "run_main.sh", remember to adjust the OpenBLAS path in "LD_LIBRARY_PATH" and the number of computing threads in "OMP_NUM_THREADS" accordingly. 

The high-dimensional data "V" is a "n" by "d" matrix, where "n" is the number of samples, "d" is the number of features. The MIM algorithm extracts the low-dimensional projection basis "S" in the following way:

    S = [];
    for i = 1 : num_tsf_dim
        S_i = f(max_dim); % f(max_dim) returns a "d" by "max_dim" orthonormal basis "S_i"
        S = [S S_i];
    end

This produces a "d" by "num_tsf_dim * max_dim" orthonormal basis "S". The low-dimensional data "U" is then computed by U=V*S
    

The "run_main.sh" takes the 7 arguments:
    1) train_data:    the labeled training data saved as a "n" by "d" matrix in an ASCII format file. Each row corresponds to a sample, each column corresponds to a feature. The samples belonging to the same class should be grouped together.
    2) test_Data:     the unlabeled test data in the same format as train_data.
    3) opt:           the option file specifying the value of "num_tsf_dim"
    4) lab_train:     the corresponding training labels saved as a "n" by "1" matrix, should start from "1" end with "C" - the number of classes. Since the samples of the same class are together, the lab_train should look like this [1 1 1 2 2 3 3 3 3 ... C C C]^T.
    5) train_red:     the projected low-dimensional training data.
    6) test_red:      the projected low-dimensional test data.
    7) tsf:           the "d" by "num_tsf_dim * max_dim" projection orthonormal basis "S".

******************
* Control option *
******************

The MIM algorithm is controlled by the two option files "main_options" and "dfe_options":

"main_options": 
    1) num_tsf_dim:    an integer >=1
    
"dfe_options":
    1) max_eigs_ite:   the maximum number of iterations the "eigs" function in the "Spectra" library could take to compute the eigenvalues, eigenvectors
    2) eigs_tol:       the convergence criterion of the "eigs" function
    3) max_dim:        the number of eigenvalues to be computed by the "eigs" function
    4) read_initial:   set to 0, do not initialize the rho value.
    5) centralize:     if set to 1, center the data according to train_data.
    6) normalize:      if set to 1, normalize the data according to train_data.
    7) eta:            the step size eta in the original paper
    8) max_ite:        the maximum number of iterations on the MIM algorithm
    9) nr_fold:        the number of folds in the cross validation, every class should have sufficient samples so that in each fold, the number of samples per class should be >2. nr_fold should be >=2.
    10) num_precision: added to avoid log 0.
    
*******************
* S-MIM and C-MIM *
*******************    

Let C be the number of classes.

"S-MIM" will extract the projection dimensions one by one.
    1) please set max_dim in "dfe_options" to 1
    2) please set num_tsf_dim in "main_options" to the maximum number of dimensions you want to extract. The recommended value should be greater than C.

"C-MIM" will extract the projection dimensions in one batch.
    1) please set max_dim in "dfe_options" to the number of dimensions you want to extract. the recommended value is C-1.
    2) please set num_tsf_dim in "main_options" to 1.
    
***********
* Example *
***********

1) Download the accompanying data files from Shuai's homepage and use the included matlab script to write the data into ASCII files.
2) Copy the generated ASCII files to the current directory.
3) Run the following command in the terminal, remember to set the path in run_main.sh correctly

    ./run_main.sh ./train_101_1 ./test_101_1 ./options ./train_lab_1 ./train_101_1_dfe ./test_101_1_dfe ./tsf_101_dfe
    
You are welcome to try the MIM algorithm on your own data, don't forget to change the parameters accordingly! Thanks:)