Tutorial for Tuner Subclassing

[omalleyt12]

No description provided.

[fchollet]

Thank you for the PR!

[fchollet]

This is a scalar so prefer doing float(run_train_step(data)) (more readable than .numpy(), and gives you a plain Python float).

[omalleyt12]

Done

[fchollet]

It should be specified in the text above ("Typically, Tuner.run_trial is the only method that users need to override when subclassing Tuner.") that run_trial should call self.on_epoch_end and optionally other callback methods such as on_epoch_begin, etc. The list of available methods, and which ones are optional or not, should be made explicit.

[omalleyt12]

Done

[fchollet]

Besides the detailed example below, I would recommend having a sketch of the code here, e.g.

class MyTuner(kt.Tuner):

    def run_trial(self, trial, ...):
        hp = trial.hyperparameters
        model = self.hypermodel.build(trial.hyperparameters)
        for epoch in range(10):
              epoch_loss = ...
              self.on_epoch_end(trial, model, epoch, logs={'loss': epoch_loss})

Otherwise it's hard to parse what's essential and what's not.

[omalleyt12]

Done

[fchollet]

This is a complex example. We should consider using a format that enables us to make the example directly runnable in Colab in one click, and easily testable in an automatically generated integration test.

For instance, examples on tensorflow.org are typically shipped as runnable notebooks. The notebooks are used to generated a HTML webpage (it goes through markdown first), can be launched in Colab, and are run as part of integration tests.

[omalleyt12]

Is there a way to do the equivalent on GH? I couldn't find anything from googling around a bit

I ran this example on my computer locally to make sure it worked, is that ok for now?

[fchollet]

There is no ready-made solution that I'm aware of, we'd have to roll out our own (and make it shared cross all of our repos). We'll do this eventually, but we can't block tutorials based on this, so let's ship the tutorial as is for now.

[omalleyt12]

Sounds good, made a tracking bug here: keras-team/keras-autodoc#35

[omalleyt12]

Thanks for the review!