Training as listener

[jSaso]

Description

Implement training as listener, so we can subscribe to training listener. We need more information to each of the provided listeners:

• normalize/de-normalize class (separate feature request)
• train data set
• test data set
• last best model
• last best model score

In this way we can do whatever we want:

• calculate prediction from best model data set with our test data set,
• de-normalize test data set and compare with de-normalized predicted data set
• create our own evaluation with real vs. predicted data
• create evaluation on train data
• create evaluation on test data
• on complete we want latest best model (with best score), the sam onEpoch listener
• on completion listener we can save training model for later use - transfer learning
• combined with early stop configuration - so we can check which termination condition stop training (separate feature request)

Will this change the current api? How?

Yes, training should be implemented as listener, so we can subscribe to listener.
Interface should have:
onStart
onEpoch
onCompletion

Who will benefit from this feature?

Everybody, subscribe to the listener and we have all information we need

References

Reference implementation: https://github.com/eclipse/deeplearning4j/blob/b5f0ec072f3fd0da566e32f82c0e43ca36553f39/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/earlystopping/listener/EarlyStoppingListener.java
but we need more information inside each of onStart, onEpoch and onCompletion method - see above

[keerthanvasist]

Thank you for creating this issue @jSaso.

[zachgk]

@jSaso We have a TrainingListener class that was added in 0.3.0. Does this look like it will work or do you see any issues with it?

[zachgk]

Since we have this feature and there is no response, I will close the issue for now. Please feel free to reopen or open a new issue