Adaptation on new dataset

[JRMeyer]

>>> prabhudatta.das09
[January 23, 2018, 9:59am]

I was wondering how to the use the best checkpoint model on new dataset
slash [let's say 5 hours of new data slash ] quickly without doing a full epoch
training. Freezing inner layers and just using FC doesn't help in
learning. Is there any other way - just like Kaldi adapts

[This is an archived TTS discussion thread from discourse.mozilla.org/t/adaptation-on-new-dataset]

[JRMeyer]

>>> reuben
[January 23, 2018, 10:11am]



> Freezing inner layers and just using FC doesn't help in learning.

In that case, you're probably stuck with fine tuning the whole network.
An epoch should be pretty fast on just 5 hours of data though.

[Archived Post]

[JRMeyer]

>>> prabhudatta.das09
[January 23, 2018, 11:52am]

My model has been trained on 2000hours of data. If I add more 30 hrs and
run epoch from checkpoint , the loss shoots up which is obvious. Is
there any intuitive paper on how to fine tune the whole network?

[Archived Post]

[JRMeyer]

>>> prabhudatta.das09
[January 24, 2018, 7:52am]

with any sort of fine tuning
like changing lr , the model overfits to the new dataset while
performing worse on the overall test set . Any suggestions?

[Archived Post]