Monte Carlo (MC) dropout in Keras with R

Question

How to implement Monte Carlo dropout with Keras in Convolutional neural networks to estimate predictive uncertainty as suggested by YARIN GAL? I am using R.R-Code is here

I am fitting the model in small batches and want to evaluate the model in small batches as well with Monte Carlo dropout.Could not find any hint in Keras documentation.BTW, I trained my model with flag training=TRUE.

Thanks

Anyone has a pythonic implementation for this one? – Arka Mallick Dec 14 '18 at 02:24 — Arka Mallick, Dec 14 '18 at 02:24
@Haramoz did you see my answer? :) – MiniQuark Feb 22 '19 at 04:07 — MiniQuark, Feb 22 '19 at 04:07
yes I realized later and upvoted already. Thanks :) – Arka Mallick Feb 22 '19 at 12:09 — Arka Mallick, Feb 22 '19 at 12:09

MiniQuark · Accepted Answer · 2018-08-26T15:35:29.933

Regular dropout only drops neurons randomly at training time, not a test time, so this is the default behavior of the Dropout class. If you want MC dropout, you need to use training=TRUE at test time as well, and you must run the forward pass multiple times: this will give you a distribution of predictions, which you can use as you please, for example to compute the mean.

I'm not familiar enough with R, so here is the class I use instead of the standard Dropout class:

class MCDropout(keras.layers.Dropout):
    def call(self, inputs, training=None):
        return super(MCDropout, self).call(inputs, training=True)

Monte Carlo (MC) dropout in Keras with R

1 Answers1