Errors while running example files

[Humbulani1234]

Describe the bug

I am experiencing the following error message while trying to run the example file autokeras.examples.automodel_with_cnn.py:

2024-11-20 20:24:34.105906: W tensorflow/core/framework/local_rendezvous.cc:404] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence
> /home/humbulani/tensorflow/env/lib/python3.12/site-packages/autokeras/auto_model.py(391)_analyze_data()
-> for analyser in analysers:
(Pdb) c
Traceback (most recent call last):
  File "/home/humbulani/tensorflow/examples_auto/automodel_with_cnn.py", line 39, in <module>
    auto_model.fit(x_train, y_train, epochs=1)
  File "/home/humbulani/tensorflow/env/lib/python3.12/site-packages/autokeras/auto_model.py", line 296, in fit
    self._analyze_data(dataset)
  File "/home/humbulani/tensorflow/env/lib/python3.12/site-packages/autokeras/auto_model.py", line 392, in _analyze_data
    analyser.finalize()
  File "/home/humbulani/tensorflow/env/lib/python3.12/site-packages/autokeras/analysers/output_analysers.py", line 80, in finalize
    raise ValueError(
ValueError: Expect the target data for classification_head_1 to have shape [2], but got [].

Attempted fixes

I read the file automodel_with_cnn.py and I couldn't find any errors, AutoKeras API usage seems to ok together with functions and classes signatures - so, I don't think the error is on API usage.

I decided to read AutoKeras source code to trace the error with pdb. The error is from the following code snippet from the function ClassificationAnalyser.finalize:

...
# Check shape equals expected shape.
        expected = self.get_expected_shape()
        actual = self.shape[1:]
        if len(actual) == 0:
            actual = [1]
        if self.encoded and actual != expected:
            raise ValueError(
                "Expect the target data for {name} to have "
                "shape {expected}, but got {actual}.".format(
                    name=self.name, expected=expected, actual=self.shape[1:]
                )
            )

The above code snippet calls the function 'get_expected_shape' reproduced below:

def get_expected_shape(self):
        # Compute expected shape from num_classes.
        if self.num_classes == 2 and not self.multi_label: 
            return [1]
        return [self.num_classes]

The 'actual local variable is not equal to the expected local variable, so it raises a ValueError exception. I Then decided to force them to be equal by removing not keyword from the line: if self.num_classes == 2 and not self.multi_label:. And it worked. I don't know yet implications of this on the whole codebase but I hope I proper fix can be found.

One implication of the above fix is that we have to include the keyword argument multi_label=True in the instantiation of the classifier: clf = ak.TextClassifier(max_trials=3, multi_label=True) from the example file imdb.py.

Note There are other small issues like numpy dtypes deprecation where we dtypes like str has changed to str_ from example file like imdb.py. I suspect this is probably due to numpy versions.

Tests

I ran only the tests from the analysers directory with no errors.

Environment

python == 3.12
autokeras == 2.0.0
keras == 3.6.0
tensorflow == 2.16.2

To reproduce

Please run the example files from the autokeras/examples directory.