Final fit not replacing the model in training history

[twprinston]

Hi,

if I run clf.final_fit(x_train, y_train, x_test, y_test, retrain=True)
then call
model = clf.load_searcher().load_best_model().produce_keras_model()
Is this model the retrained model?

I traced the 0.2.18 ImageSupervised.final_fit() source code but I'm not sure it's true.
load_best_model() seems to return the original best model in searching phase.

Thank you.

[haifeng-jin]

This might be a bug. I will investigate it. @twprinston Thanks.

[kavehtp]

Faced a similar problem:

After training the model using clf.fit() function, calling clf.final_fit() does not make any difference. The best model remains the same and does not get updated. So:

clf = ak.ImageClassifier(verbose=True, resume=True, path='autokeras-run')	
clf.final_fit(x_train, y_train, x_dev, y_dev, retrain=True)
y_test_pred = clf.predict(x_test)

leads to exactly the same results as:

clf = ak.ImageClassifier(verbose=True, resume=True, path='autokeras-run')
y_test_pred = clf.predict(x_test)