Seeking urgent help in loading a saved model trained in earlier Autokeras version

[dsbanerjee90]

Hi,

We are facing issues in loading a model which was trained with an earlier version of Autokeras.

We trained our model using, 1.1.0 (most possibly otherwise 1.0.20)
Tensorflow TF: 2.8
Python: 3.8/3.9

We tarined our model like this,

"import autokeras as ak
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X_t_v, y_t_v, test_size=0.2)#, random_state=42)
regressor = ak.StructuredDataRegressor(
    overwrite=True,
    max_trials=200  # Number of different models to try
)

"

We saved our model like this,

model.save("TRAINED/best_model_autokeras", save_format='tf')

in best_model_autokeras we have these files,
assets fingerprint.pb keras_metadata.pb saved_model.pb variables

And we are now trying load our model like this,

from keras.models import load_model
import autokeras as ak

model_newscaled = load_model('TRAINED/best_model_autokeras/')

showing this error,

"---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[3], line 1
----> 1 model_newscaled = load_model('best_model_autokeras/')

File ~/miniconda3/envs/autokeras3/lib/python3.8/site-packages/keras/utils/traceback_utils.py:67, in filter_traceback.<locals>.error_handler(*args, **kwargs)
     65 except Exception as e:  # pylint: disable=broad-except
     66   filtered_tb = _process_traceback_frames(e.__traceback__)
---> 67   raise e.with_traceback(filtered_tb) from None
     68 finally:
     69   del filtered_tb

File ~/miniconda3/envs/autokeras3/lib/python3.8/site-packages/keras/saving/saved_model/load.py:532, in KerasObjectLoader._revive_layer_or_model_from_config(self, metadata, node_id)
    530 builtin_layer = layers_module.get_builtin_layer(class_name)
    531 if builtin_layer:
--> 532   raise RuntimeError(
    533       f'Unable to restore object of class \'{class_name}\' likely due to '
    534       f'name conflict with built-in Keras class \'{builtin_layer}\'. To '
    535       'override the built-in Keras definition of the object, decorate '
    536       'your class with `@keras.utils.register_keras_serializable` and '
    537       'include that file in your program, or pass your class in a '
    538       '`keras.utils.CustomObjectScope` that wraps this load call.') from e
    539 else:
    540   raise

RuntimeError: Unable to restore object of class 'Normalization' likely due to name conflict with built-in Keras class '<class 'keras.layers.preprocessing.normalization.Normalization'>'. To override the built-in Keras definition of the object, decorate your class with `@keras.utils.register_keras_serializable` and include that file in your program, or pass your class in a `keras.utils.CustomObjectScope` that wraps this load call.
"

Can you please help me with this issue? We have an article that is under review, and we need to rerun an experiment but we are struggling with this issue. Any help would be much appreciated. Thank you.

Best regards,
Deep Banerjee
Plymouth Marine Laboratory, UK

[vikasprajapati-01]

Hi @dsbanerjee90 ,
I think the error occurs because the model was trained with AutoKeras 1.0.20 / 1.1.0 and TensorFlow 2.8 (Python 3.8/3.9), but it’s now being loaded in a newer Keras/TensorFlow setup. In newer versions, Keras has its own built-in Normalization layer, which clashes with how AutoKeras 1.x saved that same layer. Because of this version mismatch, Keras gets confused while loading the model and throws the error, even though the model files themselves are fine.

In my opinion, the safest and simplest fix is to recreate the original training environment. Use Python 3.8, TensorFlow 2.8.0, and AutoKeras 1.0.20 or 1.1.0, and avoid installing the standalone keras package. Then load the model using tf.keras.models.load_model (not keras.models.load_model) and pass AutoKeras custom objects if needed.
This might resolve the issue.