I’ve noticed a phenomenon wherein the EON Tuner will produce a model with a high accuracy score, but with the Data Augmentation option turned off. Once retraining with DA enabled, very often the score will drop dramatically as the model may not be resilient to noise. Would it be possible to specify DA as an option to the Tuner so this might be taken into consideration when computing the score?
Hi @jefffhaynes, yes, we’re adding this in a next iteration (incl. full control of all parameters that the tuner works through).
Indeed this is an idea we have been considering. I’m going to fast-track this and it’ll be live before the end of this week. As soon it’s live i’ll post an update here.
Full control of the EON Tuner hyperparameter search space will follow in the coming 1-3 months.
Data augmentation for the EON tuner is live now. On the EON Tuner results page you’ll see ‘Data augmentation’ as first layer in the ‘Keras’ on the model cards for models for which data augmentation is enabled.
Currently we are always using the ‘Add noise’ and ‘Mask time bands’ settings set to ‘low’ in the EON Tuner. In the future we’ll be extending this further and add support to vary the data augmentation settings.
I don’t know if this behavior changed but at least for MFE models, the data augmentation option in the Tuner no longer appears to be turned on by default.
EDIT: Actually as far as I can tell, it’s disabled for both models. At least, when I switch over to the NN model page it isn’t enabled.