An observer that provides for early-stopping and saves the minima of models to file.
Could also publish to channel "test" to tell the tester to test the model. But this could be done at the end by unserializing the minima of the model, and testing it. For now, just test every time "test" is advised. This will however make for holes in the learning curves of the test set... The payoff is faster convergence since less time spent on testing. This could be an option of the experiment.
An observer that provides for early-stopping and saves the minima of models to file. Could also publish to channel "test" to tell the tester to test the model. But this could be done at the end by unserializing the minima of the model, and testing it. For now, just test every time "test" is advised. This will however make for holes in the learning curves of the test set... The payoff is faster convergence since less time spent on testing. This could be an option of the experiment.