The value of energy regression by DNN method is negative

Dawei · August 30, 2022, 1:40pm

I use DNN method to reconstruct gamma energy, the energy range of the training set is 20-50MeV, but the energy predictions are negative as shown below. However, the activation function of the output layer of my neural network is RELU, this confuses me.

Dawei · August 30, 2022, 1:43pm

Could you help me？@moneta,I am sorry to bother you again.

moneta · August 30, 2022, 2:06pm

Hi,
Can you please share the code and input data, so I can have a closer look ?
Thanks

Lorenzo

Dawei · August 30, 2022, 2:32pm

Hi, thanks for your reply! This my code and data.Because of the 3MB limit of the file size, I remove some data of the root file
Thanks,
David

DNNRegression.C (3.3 KB)
TrainData.root (2.7 MB)

moneta · August 31, 2022, 9:38am

Hi,

The problem is in using the old kDNN method that is now deprecated. You should use the new kDL method. The correct booking string is:

factory->BookMethod(dataloader, TMVA::Types::kDL, "DNN_CPU", nnOptions);

By doing this I think I am getting only positive results and a much better regression.

Cheers
Lorenzo

Dawei · September 1, 2022, 6:06am

Hi, moneta,Thank you a lot. By changing “KDL”, I have got a better regression