Compatibility Issues with tf.keras.metrics.F1Score on Python 3.9 and 3.10

[rizoudal]

With the recent update where TensorFlow Addons deprecated support for F1Score and by using the F1Score function of tf.keras, problems arise. Specifically, the num_classes parameter is no longer supported in the tf.keras.metrics.F1Score.

After updating the code to import F1Score from tf.keras and removing the num_classes parameter, the following errors occur during testing on GitHub Actions with Python versions 3.9 and 3.10:

ValueError: Tensor conversion requested dtype float32 for Tensor with dtype int64: <tf.Tensor 'IteratorGetNext:1' shape=(None, None) dtype=int64>

and/or

ValueError: Tensor conversion requested dtype float32 for Tensor with dtype uint8: <tf.Tensor 'IteratorGetNext:1' shape=(None, None) dtype=uint8>

Errors occur in the following tests:

• tests/test_automl_block_train.py
• tests/test_automl_block_pred.py

Please review and address this issue to maintain compatibility with the new tf.keras metric implementations.

Thank you!

[muellerdo]

Maybe the following could be a solution:

in data_processing/data_generator.py

in line 296:

batch_stack[1].extend(np.float32(self.labels[index_array]))

add np.float32() cast to labels