CvEM::read_params() read wrongly? (Bug #1340)

Hi, all:

Finally, I noticed that this is probably a bug of CvEM, because it's using dual systems to record the parameters. A CvEMParams already contains all parameter matrices. Please refer to what's the class definition in current CvEM and CvEMParams.

The following 6 lines are in CvEM

CvEMParams params;
....
    CvMat* means;
    CvMat** covs;
    CvMat* weights;
    CvMat* probs;

And CvEMParams is defined as:

struct CV_EXPORTS_W_MAP CvEMParams {
CvEMParams();
CvEMParams( int nclusters, int cov_mat_type=1/*CvEM::COV_MAT_DIAGONAL*/,
int start_step=0/*CvEM::START_AUTO_STEP*/,
CvTermCriteria term_crit=cvTermCriteria(CV_TERMCRIT_ITER+CV_TERMCRIT_EPS, 100, FLT_EPSILON),
const CvMat* probs=0, const CvMat* weights=0, const CvMat* means=0, const CvMat** covs=0 );

CV_PROP_RW int nclusters;
    CV_PROP_RW int cov_mat_type;
    CV_PROP_RW int start_step;
    const CvMat* probs;
    const CvMat* weights;
    const CvMat* means;
    const CvMat** covs;
    CV_PROP_RW CvTermCriteria term_crit;
};

The reason why I failed to read_params() after write_params() is
when writing, the trained data were stored in matrices (CvMat* means;
CvMat** covs; CvMat* weights; CvMat* probs;) declared in CvEM;
while reading, the trained data are now loaded into the parameter CvEMParams params.

I'm guessing this is most probably a bug. Waiting for the fixing.

Best Regards
Pei

Replying to jiapei100:

Hi, all:

It looks there is some problem with CvEM::write_params() and CvEM::read_params() .

Whenever I didn't "read after write" the CvEM parameters, the predict() function works fine. However, whenever I called read_params() after I write_params() of a CvEM model, the predict() function goes crash.

I called the function in this way:
gmmModel.read_params(fs, 0);

Then, I noticed in the documentation of CvEM::read_params(),

If it is NULL, the function searches a node with parameters in all the top-level nodes (streams), starting with the first one.
Does that mean that there is no simple way to load all parameters in all level modes, instead of just the top-level nodes ????

Anyway, my trained .xml file is copied in the following and attached as file color001.xml:
please do give me a hand on how to CvEM::read_params() successfully with all level parameters loaded using CvEM::read_params() ???

Best Regards
Pei

<opencv_storage>
<params>
<cov_mat_type>COV_MAT_DIAGONAL</cov_mat_type>
<start_step>START_AUTO_STEP</start_step>
<nclusters>3</nclusters>
<weights type_id="opencv-matrix">
<rows>1</rows>
<cols>3</cols>
<dt>d</dt>
<data>
4.7505367397911979e-001 4.0513001653020758e-001
1.1981630949067631e-001</data></weights>
<means type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
4.6735226574391647e-001 4.3428615328431247e-001
7.7229067746570723e-001 2.8662794138474235e-001
2.4997266948677280e-001 5.2728525926602976e-001 1. 1. 1.</data></means>
<covs>
<_ type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
3.1503335914751878e-003 0. 0. 0. 3.2953545487618106e-003 0. 0.
0. 7.2003084647408845e-003</data></_>
<_ type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
9.8828026607258000e-004 0. 0. 0. 1.3064142972168662e-003 0. 0.
0. 2.6975317757247349e-003</data></_>
<_ type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
1.1920928955078125e-007 0. 0. 0. 1.1920928955078125e-007 0. 0.
0. 1.1920928955078125e-007</data></_></covs></params>
</opencv_storage>

Problem is now bypassed by using CvStatModel::save() and CvStatModel::load() .

But, personally speaking, that doesn't mean CvEM::write_params() and CvEM::read_params() are a functionally corresponding pair.
Apparently, they are not.

Still think this is a bug.
Looking forward to the fixing.

Best Regards
Pei

Replying to jiapei100:

Hi, all:

It looks there is some problem with CvEM::write_params() and CvEM::read_params() .

Whenever I didn't "read after write" the CvEM parameters, the predict() function works fine. However, whenever I called read_params() after I write_params() of a CvEM model, the predict() function goes crash.

I called the function in this way:
gmmModel.read_params(fs, 0);

Then, I noticed in the documentation of CvEM::read_params(),

If it is NULL, the function searches a node with parameters in all the top-level nodes (streams), starting with the first one.
Does that mean that there is no simple way to load all parameters in all level modes, instead of just the top-level nodes ????

Anyway, my trained .xml file is copied in the following and attached as file color001.xml:
please do give me a hand on how to CvEM::read_params() successfully with all level parameters loaded using CvEM::read_params() ???

Best Regards
Pei

<opencv_storage>
<params>
<cov_mat_type>COV_MAT_DIAGONAL</cov_mat_type>
<start_step>START_AUTO_STEP</start_step>
<nclusters>3</nclusters>
<weights type_id="opencv-matrix">
<rows>1</rows>
<cols>3</cols>
<dt>d</dt>
<data>
4.7505367397911979e-001 4.0513001653020758e-001
1.1981630949067631e-001</data></weights>
<means type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
4.6735226574391647e-001 4.3428615328431247e-001
7.7229067746570723e-001 2.8662794138474235e-001
2.4997266948677280e-001 5.2728525926602976e-001 1. 1. 1.</data></means>
<covs>
<_ type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
3.1503335914751878e-003 0. 0. 0. 3.2953545487618106e-003 0. 0.
0. 7.2003084647408845e-003</data></_>
<_ type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
9.8828026607258000e-004 0. 0. 0. 1.3064142972168662e-003 0. 0.
0. 2.6975317757247349e-003</data></_>
<_ type_id="opencv-matrix">
<rows>3</rows>
<cols>3</cols>
<dt>d</dt>
<data>
1.1920928955078125e-007 0. 0. 0. 1.1920928955078125e-007 0. 0.
0. 1.1920928955078125e-007</data></_></covs></params>
</opencv_storage>

looks like it's normal bug. Lowering the priority to major.

Thank you for the report. CvEM write method had a bug, it saved all matrices twice. It was fixed, but be careful due to we had to change the storing format to fix it.
Note, we added new c++ implementation cv::EM and moved CvEM to the legacy module (cv::EM will be available in coming opencv2.4). So you should change include file from ml to legacy module. The implementation of CvEM methods was replaced by the call of cv::EM ones. CvEM writes the trained model in new format now (as mentioned above).
cv::EM should solve many issues of the previous implementation. We added several new tests to check this. If you'll have some other problems from r8075, please open an issue.