Double dropouts in EON Tuner

Report bugs
#1

Hello!

I am running EON Tuner and in some configurations, after the feature extractor (convnet), there are two consecutive dropout-layers with 0.5 rates. I was wondering if this behaviour is expected, what are the benefits of this?

Thanks!

#2

Hello @ympyra,

I am not exactly sure if I would have tried that if I had to design an architecture by myself.
The EON tuner tests different possibilites and give you the best impulse corresponding to your settings.
I am asking our ML team if it has any known benefits to add two consecutive dropouts after a convnet.

Regards,

Louis

#3

I think this is a bug. I will investigate this behavior further and post an update here as soon as I know more.

Can you share your project ID and the title of a model variant for which this happens?

#4

Hello @mathijs!

Project ID is 80416 and in EON Tuner, e.g.

spectr-conv1d-bc1
mfe-conv1d-9f5
mfcc-conv1d-7ed

have double dropouts.

Thanks!

#5

Thanks @ympyra. I was able to identify the issue as a regression. I will deploy a fix to production by tomorrow at latest and post an update here as soon as the fix is live.

#6

Hello @mathijs,

Thanks for the prompt reply and check!

Great! :slight_smile:

#7

Hi @ympyra , a fix for this issue has been deployed! Apologies for the delay, resolving this issue proved to be slightly more complex than initially anticipated.

1 Like