[bug] TextVectorization + Sequential model doesn't work

[WeichenXu123]

Tensorflow version:
2.19.0-dev20241108

Keras version:
3.7.0.dev2024111103

Installation command: pip install --pre tf-nightly

Reproducing code:

import numpy as np
import tensorflow as tf


def get_text_vec_model(train_samples):
    from tensorflow.keras.layers import TextVectorization
    VOCAB_SIZE = 10
    SEQUENCE_LENGTH = 16
    EMBEDDING_DIM = 16
    vectorizer_layer = TextVectorization(
        max_tokens=VOCAB_SIZE,
        output_mode="int",
        output_sequence_length=SEQUENCE_LENGTH,
    )
    vectorizer_layer.adapt(train_samples)
    model = tf.keras.Sequential(
        [
            vectorizer_layer,
            tf.keras.layers.Embedding(
                VOCAB_SIZE,
                EMBEDDING_DIM,
                name="embedding",
                mask_zero=True,
            ),
            tf.keras.layers.GlobalAveragePooling1D(),
            tf.keras.layers.Dense(16, activation="relu"),
            tf.keras.layers.Dense(1, activation="tanh"),
        ]
    )
    model.compile(optimizer="adam", loss="mse", metrics=["mae"])
    return model


train_samples = np.array(["this is an example", "another example"], dtype=object)
train_labels = np.array([0.4, 0.2])
model = get_text_vec_model(train_samples)


# Error: ValueError: Invalid dtype: object 
model.fit(train_samples, train_labels, epochs=1)

Error stack:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/weichen.xu/miniconda3/envs/mlflow/lib/python3.9/site-packages/keras/src/utils/traceback_utils.py", line 122, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/home/weichen.xu/miniconda3/envs/mlflow/lib/python3.9/site-packages/optree/ops.py", line 747, in tree_map
    return treespec.unflatten(map(func, *flat_args))
ValueError: Invalid dtype: object

The same code works in "keras==3.6.0"

[fchollet]

It seems we're no longer detecting object arrays as string arrays, probably because we've upgraded our numpy dependency. Object arrays are ambiguous since they can contain anything, not just strings.

I recommend instead using tf.string tensors, which are explicitly strings and are also much more memory efficient:

train_samples = tf.convert_to_tensor(["this is an example", "another example"])

This would fix your code example.

[fchollet]

I fixed it at HEAD, regardless.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No