search

felixbur / nkululeko

Machine learning speaker characteristics
MIT License
31 stars 5 forks source link

Montecarlo dropout #137

Open felixbur opened 1 month ago

felixbur commented 1 month ago

nkululeko already knows the concept of "runs", i.e. running the same experiment n times with different random initializations. Would be interesting to output the ensemble performance of all these models.

bagustris commented 1 month ago

If I understand correctly, it is similar to an ensemble of different seeds. Last time, I edit svm and balancing to use the same seed so the output will be consistent for every run.

If so, the ini file is the same for different results, hence we can use the output prediction CSV file to make ensemble?

felixbur commented 1 month ago

yes, it's very similar, the variation comes from the random drop outs. I'd like to look into it, if it could be just an additional flag in the ini file