AttributeError: 'SymbolicTensor' object has no attribute '_datatype_enum'

[innat]

Works in Keras 2 but not in Keras 3.

!pip install keras-nightly -q

import tensorflow as tf # 2.15
import keras
from keras import layers
import numpy as np
keras.__version__ # 3.0.3.dev2024011803

processing_model = keras.Sequential(
    [
        layers.Normalization(
            mean=[123.675, 116.28, 103.53],
            variance=[np.square(58.395), np.square(57.12), np.square(57.375)]
        )
    ]
)

dataset = tf.data.Dataset.from_tensor_slices(
    (
        tf.random.uniform([4, 32, 224, 224, 3]),
        tf.random.uniform([4, 3]),
    )
)

for i in range(3):
    ds = dataset
    ds = ds.map(lambda x, y: (processing_model(x), y))
    print(i)

0
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
[<ipython-input-10-d75ead39f56a>](https://localhost:8080/#) in <cell line: 1>()
      1 for i in range(3):
      2     ds = dataset
----> 3     ds = ds.map(lambda x, y: (processing_model(x), y))
      4     print(i)

7 frames
[/usr/local/lib/python3.10/dist-packages/tensorflow/python/framework/tensor.py](https://localhost:8080/#) in __getattr__(self, name)
    259         tf.experimental.numpy.experimental_enable_numpy_behavior()
    260       """)
--> 261     self.__getattribute__(name)
    262 
    263   @property

AttributeError: 'SymbolicTensor' object has no attribute '_datatype_enum'

Similar issue tensorflow/tensorflow#29931 (comment)

[grasskin]

Thank you @innat, as a temporary workaround you can call processing_model.build((None, 224, 224, 3)) before the loop. Alternatively, the issue also goes away by setting tensorflow to eager mode so it seems likely that the model build is not being triggered in graph mode from within the map call.