search

Shark-ML / Shark

The Shark Machine Leaning Library. See more:
http://shark-ml.github.io/Shark/
GNU Lesser General Public License v3.0
493 stars 130 forks source link

DeepNetworkTrainingRBM #265

Open dry11orange opened 5 years ago

dry11orange commented 5 years ago

win7 64 Shark-4.0.1 DeepNetworkTrainingRBM

I change the flag_rbm, test error(0.26, 0.06) will change a lot, And they don't have relation at all. How could this happen?

DeepNetworkTrainingRBM.txt

Ulfgard commented 5 years ago

could you tell me exactly what you a) do b) observe and c) what you expect the output to be?

dry11orange commented 5 years ago

I tried to determine the application scenarios for RBM And whether RBM has irreplaceable characteristics compared with neural network

In the test code DeepNetworkTrainingRBM I commented out the RBM part, trying to compare the training effect of neural network without RBM

And then I was surprised Cut off the connection between RBM and neural network RBM still has an influence on the training results of neural network Training error 0.26->0.06

Logically speaking The hyper-parameter configuration of neural network is not good, which may lead to high training error It should not be associated with irrelevant RBM Because of this, I am not sure whether the neural network hyper-parameter need to be adjusted