compute_word_piece_vocabulary

[42Cummer]

Describe the bug

if not isinstance(data, (list, tf.data.Dataset)): #from source code
AttributeError: 'NoneType' object has no attribute 'data'

because compute_word_piece_vocabulary function is not importing TensorFlow correctly (tf = None)(refer to source code lines 4-11 and 117)

To Reproduce

def train_word_piece(dataset, vocab_size, reserved_tokens):
    word_piece_ds = dataset.unbatch().map(lambda x, y: x)
    vocab = compute_word_piece_vocabulary(
        data = word_piece_ds.batch(1000).prefetch(2),
        vocabulary_size=vocab_size,
        reserved_tokens=reserved_tokens,
    )
    return vocab

train_word_piece(train_dataset, VOCAB_SIZE, reserved_tokens)

Code from https://keras.io/examples/nlp/fnet_classification_with_keras_hub/

Expected behavior
Function executes without error.

Additional context
Tensorflow version 2.17.1
KerasHub version 0.18.1

Maybe remove lines 4-11 in source?

[mattdangerw]

I suspect there's an issue with your tensorflow or tensorflow_text install. Can you try running this?

import tensorflow
import tensorflow_text

print(tensorflow.__version__)
print(tensorflow_text.__version__)

Lines 4-11 are to allow keras-hub to function without a tensorflow install. All our preprocessing (including this function) uses tf.data and tensorflow-text (even on jax and torch), so this feature is really only for power users who want to run all preprocessing in a separate python environment from training/inference.

In your case, it's most likely an issue with your tensorflow or tensorflow-text installation.

[mattdangerw]

There is a valid bug here though. For other library symbols that need tensorflow text we print a helpful error message if the library is missing or otherwise failing to import. We are missing that for compute_word_piece_vocabulary. I'll add it now!

Thanks

[42Cummer]

Oh I see now, thanks for clarifying! But that error message would be helpful!