Training aways converges after 2 or 3 epochs about basset

[davek44]

Hi Gabriel,

Sorry I missed your comment here. My Harvard email turned off, and I just realized I wasn't getting notifications.

If you're willing to put some time in to set it up, Spearmint (https://github.com/HIPS/Spearmint) is a great option for performing hyper-parameter searches. It sounds like your models are training quickly, so maybe you can perform a grid search in a reasonable amount of time, too.

Best,
Dave

[suragnair]

@GabrielHoffman were you able to perform a search?

@davek44 are the parameters specified in https://github.com/davek44/Basset/blob/master/data/models/pretrained_params.txt the ones you used to train the final model the numbers of which you reported in the paper?

[davek44]

@suragnair Yes, I'm pretty sure that's true. I can check if you need a definitive answer. In general, it's pretty robust to small changes to the model architecture, and more sensitive to optimization parameters.

[GabrielHoffman]

Hi @suragnair, I used the default model architecture in my current application, but I hope to get all to a more thorough search soon