how do you implement BCNN？

Riashat / Active-Learning-Bayesian-Convolutional-Neural-Networks

Active Learning on Image Data using Bayesian ConvNets

135 stars 46 forks source link

how do you implement BCNN？ #2

Open ShellingFord221 opened 5 years ago

ShellingFord221 commented 5 years ago

In file Active-Learning-Bayesian-Convolutional-Neural-Networks/ConvNets/active_learning/BCNN_cifar10.py , the architecture of the model is still a CNN, rather than a BCNN. Besides, the training method in this file is SGD + momentum, which is also related to CNN other than BCNN (the training method for BCNN should be BBB for example), so how do you implement BCNN by Keras in your experiment?

zxj32 commented 5 years ago

I think it is the author using the technical "MC dropout" to do the Bayesian inference, which means the if we add dropout after each convolutional layer, then the CNN model is going to be a BCNN. Training this BCNN (minimize the loss, cross entropy for example) has the same effect to minimize the KL-divergence (variational inference) so that we can get an approximate posterior distribution.

ibarrien commented 4 years ago

any updates from the authors on this issue? i.e. how BCCN is implemented?