search

kumar-shridhar / PyTorch-BayesianCNN

Bayesian Convolutional Neural Network with Variational Inference based on Bayes by Backprop in PyTorch.
MIT License
1.42k stars 323 forks source link

num_ens #53

Closed ghost closed 4 years ago

ghost commented 4 years ago

Hi, I'm trying to apply you framework to a regression problem. in your default configuration file the num_ens variable is set to 1. I was reasoning about its influence in the training, since it is the number of samples to estimate the objective function.

Which is the optimal value for this parameter? I couldn't find any discussion about it in the paper! Thanks for help!

kumar-shridhar commented 4 years ago

Hi @Valensic,

We got good results for num_ens in the range of 5-20. 10 is a good number to start with. Not much difference when the values are higher.

ghost commented 4 years ago

Thank you