Replace keras_tuner.engine.tuner to keras_tuner.Tuner

[AlekseyGur]

Hello!

It seems that method keras_tuner.engine.tuner was replaced/renamed by keras_tuner.Tuner.
Now it's necessary to make some changes in this file: /autokeras/engine/tuner.py

Row № 30:
replace "class AutoTuner(keras_tuner.engine.tuner):" to "class AutoTuner(keras_tuner.Tuner):"

Am I right?

[AlekseyGur]

• now during model.export_model() an error accrued in same file: /autokeras/engine/tuner.py:

Error: AttributeError: module 'keras_tuner.engine.tuner' has no attribute 'maybe_distribute'

Row № 62:
with keras_tuner.engine.tuner.maybe_distribute(self.distribution_strategy):
should be replaced with:
with keras_tuner.src.engine.tuner.maybe_distribute(self.distribution_strategy):

[haifeng-jin]

I have just updated KerasTuner to 1.4.3, which allows user to access from keras_tuner.engine.tuner.Tuner.
This is for backward compatibility. Just in case other packages also imported the tuner from this path.
So I just made it available.

[AlekseyGur]

Thank you!
One problem left with maybe_distribute

[AlekseyGur]

I don't understand why StructuredDataRegressor models have such a low prediction power.

My situation:
During training (using x_train and y_train) looses are quite low (MSE=50, val_loss=10). But if I try to use trained and exported model to predict x_train I get results that not even close to y_train. Can someone admit that situation?

Seems like export_model return the worst underfitted model. Results differs very much. y_train contains values [-10, ... , 60] but predicted y_train contains values [-2, ... , +2].

In my dataset there are 22 features, 1700 examples, and simple regression performs good (good prediction power).

# Training
reg = ak.StructuredDataRegressor(
    overwrite=True,
    seed=42,
)
reg.fit(
    tf.data.Dataset.from_tensor_slices((x_train.astype(float), y_train.astype(float))),
    validation_split=0.25,
    verbose=1,
)
model = reg.export_model()

# Prediction
pred = model.predict(x_train.astype(float))

I tried to use x and y during fitting but results are the same:

reg.fit(
    x=x_train.astype(float),
    y=y_train.astype(float),
    validation_split=0.25,
    verbose=1,
)

[AlekseyGur]

Oh, I,ve Just found a reason. I forgot to shuffle df before using it in a fit() method.