Is it possible to skip the correlation computation before CNN training?

TMVA
1

Hi, I’m using TMVA::DNN::CNN to train a classifier. My data is some 2×60×60 images. I found that even if I set “Transformations=I” (i.e. identity transformation without decorrelation) , it still computes the 7200×7200 correlation matrix at the beginning, which is very time-consuming. Could anyone tell me if it’s possible to skip this process before training CNN?

1 Like
2

Hi

You can add this option to the DataLoader::PrepareTrainingAndTestTree function:

!CalcCorrelations

This should skip that computations. If you have still problems please let me know
Best regards
Lorenzo

1 Like
3

That works for me. Thanks a lot! :grinning:

1 Like