search

hsieh42 / exit_process

Exit process and project boards for tracking issues.
0 stars 0 forks source link

Adapt LIBRA to MatlabR2018 #16

Open hsieh42 opened 5 years ago

hsieh42 commented 5 years ago

svnclassify and svntrain no longer there in R2018. See https://www.mathworks.com/help/stats/svmclassify.html

Sent with GitHawk

hsieh42 commented 5 years ago

/cbica/home/hsiehm/Software/LIBRA-SVN/branches/LIBRA-1.0/Source/Code/libra_train.m and /cbica/home/hsiehm/Software/LIBRA-SVN/branches/LIBRA-1.0/Source/Code/libra_run.m have been modified to work with fitcsvm and predict. Now testing the crossval function:

Mdl=fitcsvm(fullset(:,SVMmodel_vars+1),fullset(:,end), 'CrossVal','on');
genError = kfoldLoss(Mdl)

But CPU time limit exceeded (core dumped) on an interactive node.

Once everything tested, need to rerun the training for 5 models (GE/Hologic, Raw/Processed, and LIBRA_SVM).

hsieh42 commented 5 years ago

See https://cbica-infr-velog.uphs.upenn.edu/elog/PC/1558.

The accuracy after changing the functions svmtrain and svmclassify to fitsvm and predict has been drastically deteriorated.

Original R2014B (Case1.dcm, PD=74.80%, in /cbica/home/hsiehm/Software/LIBRA-SVN/branches/LIBRA-1.0_skane_one_model/Source/Demo_Test):

R2018A (PD=2.37%, in /cbica/home/hsiehm/Software/LIBRA-SVN/branches/LIBRA-1.0/Source/Demo_Test):

gastouna commented 5 years ago

We need to follow up with the CBICA admins: https://cbica-infr-velog.uphs.upenn.edu/elog/PC/1697