Please support RL and custom training loops

[bionicles]

To "pull" hyperparameters is a nice way to write less code, I really like the API for this repo, but we use a custom GradientTape training loop for multiple different multi-step tasks, model.compile/model.fit doesnt work for us

how do you use this repo for custom training loops?

[omalleyt12]

@bionicles this is definitely on the roadmap, on the Keras Tuner side as well as making it easier on the Keras side to plug custom training step into compile/fit

On the Keras Tuner side, for complicated use cases the current best way imo is to override the Tuner class and replace calls to fit with calls to your custom training logic (and make sure to report the results to the Oracle and call the callback hooks as necessary)

[bionicles]

i'd like to use keras-tuner but im unclear on this codebase and there's zero documentation on how to use the callback hooks. custom != complicated, i'll just use optuna for now

[omalleyt12]

@bionicles There will be more documentation ahead of the 1.0 launch (coming soon!). The codebase is still in active development, so some things for advanced use cases are still internal but will be exposed externally soon. For now, here's an example of how to implement a custom training loop that should remain stable:

class MyTuner(kerastuner.engine.Tuner):
  def run_trial(self, trial, x, y, val_x, val_y):
        model = self.hypermodel.build(trial.hyperparameters)
        for epoch in range(10):
          self.on_epoch_begin(trial, model, epoch, logs={})
          for batch in range(100):
             self.on_batch_begin(trial, model, batch, logs={})
             outputs = model(x)
             loss = myloss(outputs, y)
             self.on_batch_end(trial, model, batch, logs={'loss': loss})
          val_loss = ...
          self.on_epoch_end(trial, model, epochs, logs={'val_loss': val_loss})

def build_model(hp):
   num_layers = hp.Int('num_layers', 1, 10)
   ...
   return model

tuner = MyTuner(
  oracle=kerastuner.tuners.randomsearch.RandomSearchOracle(
    objective='val_loss',
    max_trials=30),
  hypermodel=build_model)
tuner.search(x, y, val_x, val_y)

[omalleyt12]

I'll update this thread again when we have documentation available for best practices for these use cases

[omalleyt12]

Closing as there is now a pending PR for a tutorial on custom training loops: #136