Autokeras: I can not use the method predict on keras model (AutoModel.export_model()) with categorical features

[leandroimail]

Bug Description

When I use the StructuredDataClassifier and I export the model by export_model, I can not use the predict method - model.predict(test_data.to_numpy()) --, because the Keras model can not handle with categorical features. I believe that export_model does not add in the pipeline model the preprocess routine to treat the categorical features.

The error presented is: "ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type float)."

below the code:

Bug Reproduction

Code for reproducing the bug:

train_features = dataset_train_features.to_numpy()
train_label = dataset_train_label.to_numpy()
test_features = dataset_test_features.to_numpy()
test_label = dataset_test_label.to_numpy()

estimator = ak.StructuredDataClassifier(**params)

estimator.fit(x=train_features,
y=train_label,
validation_data=(test_features, test_label),
epochs=epochs)

##### THIS PART WORKS

predict_values = estimator.predict(test_features)
eval_values = estimator.evaluate(test_features, test_label)

##### THIS PART DOES NOT WORK

model = estimator.export_model()
predict_values = model.predict(test_features) ####### here it is throw the error #########
predict_proba = model.predict_proba(test_features)

#

Data used by the code:
Titanic Dataset

Expected Behavior

Setup Details

Include the details about the versions of:

• OS type and version: Linux Ubuntu
• Python: 3.7
• autokeras: 1.0.4
• keras-tuner: keras-tuner-1.0.2rc0
• scikit-learn:
• numpy:
• pandas:
• tensorflow: 2.2.0

Additional context

I read the autokeras code and it only returns self.tuner.get_best_model(). I believe it is needed to add it in the same pipeline preprocessor and model.

##### export_model only get the best model but do not make a pipeline with the preprocessors

autokeras.autokeras.auto_model.py
class AutoModel(object):
def export_model(self):
return self.tuner.get_best_model()

[haifeng-jin]

The prerprocess steps are using preprocessing layers of Keras, it should be part of the model. Would you paste your model.summary()?
I guess it might be cause of the format of the training data.
We use np.unicode for the data.

[leandroimail]

Thank you @haifeng-jin for your response.
When I put np.unicode this works to me. But I believe that is important to use the original type of numpy, or perhaps, you can put this transformation within the preprocess. This is only an idea/suggestion.
Again, thank you for your help!
Bellow, I put the de model.summary()

---

model.summary()
Model: "model"

---

Layer (type) Output Shape Param #

input_1 (InputLayer) [(None, 9)] 0

---

multi_column_categorical_enc (None, 9) 0

---

dense (Dense) (None, 64) 640

---

batch_normalization (BatchNo (None, 64) 256

---

re_lu (ReLU) (None, 64) 0

---

dropout (Dropout) (None, 64) 0

---

dense_1 (Dense) (None, 16) 1040

---

batch_normalization_1 (Batch (None, 16) 64

---

re_lu_1 (ReLU) (None, 16) 0

---

dropout_1 (Dropout) (None, 16) 0

---

dense_2 (Dense) (None, 32) 544

---

batch_normalization_2 (Batch (None, 32) 128

---

re_lu_2 (ReLU) (None, 32) 0

---

dropout_2 (Dropout) (None, 32) 0

---

regression_head_1 (Dense) (None, 1) 33

Total params: 2,705
Trainable params: 2,481
Non-trainable params: 224

---

[mdturp]

Similar Issue here with the StructuredDataRegessor

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Arcyno]

Still having the bug with 1.0.12

[Arcyno]

Ok I have managed to understand the issue for me : all the numeric data have to me cast as str before calling "predict()" (though I am not sure why). I hope it will help people