Combining the outputs of multiple models with CTC output layer(ensemble)

Asked Apr 16 '19 at 06:41

Active Apr 16 '19 at 06:41

Viewed 269 times

I am trying to combine the results of different models to give a better result(like voting in classification).

Output of my models are from ctc layer, which is probability distribution of the output class for time stamp. It has 2d output (T X P) T is the time stamp and P probability Distribution of the output class.

Simple average didn't work as there are some problems with the alignment for different models, ie. Outputs can be _ _ b _ _ and b _ _ _ _ for two models.

Is there a better way to combine the outputs of different models with ctc loss?

asked Apr 16 '19 at 06:41

newlearnershiv

What about geometric instead of arithmetic mean? Combine them within beam search (i.e., work with recombined hypotheses)? – Jindřich Apr 16 '19 at 07:34
@Jindřich if we take geometric mean, it won't add up to one right? – newlearnershiv Apr 17 '19 at 08:10
Oh, of course, you're right. – Jindřich Apr 17 '19 at 13:07

Combining the outputs of multiple models with CTC output layer(ensemble)

0 Answers0