can't load model saved by tensorflow.keras.callbacks.ModelCheckpoint

[patience111]

Dear developers,

I met a problem when I use keras-nlp for building and testing a multi-label mutil-class model.
Because my model is just classification, so I only used keras_nlp.layers.TransformerEncode for building the transformer structure (as the example showed in "Quick introduction", but this brief introduction has no example of model loading when test). I want to save the model with the highest training accuracy, so I used tensorflow.keras.callbacks.ModelCheckpoint for monitering the training process and update&save the best (highest training acc) model for testing later, and save as "xxx.h5" .
At the point of testing, I used tf.keras.models.load_model('xxx.h5'), I met the error: "ValueError: Unknown layer: TransformerEncoder. Please ensure this object is passed to the custom_objects argument. See https://www.tensorflow.org/guide/keras/save_and_serialize#registering_the_custom_object for details."
I also checked the "xxx.h5" file, it only has ~400k (not as much large as my previous CNN model using the same way for saving model and also in the framework of keras in TF, in general, it has ~300M), I suspected that the "xxx.h5" just only saved the weight not the model structure & weight. And I tried to define the model first and used tf.keras.Model.load_weights for load the "xxx.h5" but also didn't work.
I searched online and couldn't find any good solution. could you please help me with this problem for saveing & loading model when the model built with keras-nlp especially I want to save the best model using checkpoint? ( or when using layers in keras-nlp, the model could only be saved by .save() method after training not work on saving highest training acc by using tensorflow.keras.callbacks.ModelCheckpoint ?)

Thank you very much!!

[patience111]

oh, maybe my TF version is not >= 2.9, I'll try upgrade first.
Thank you

[patience111]

the same error"ValueError: Unknown layer: TransformerEncoder. Please ensure this object is passed to the custom_objects argument."

[patience111]

still not okay by refering "https://www.tensorflow.org/guide/keras/save_and_serialize#registering_the_custom_object "....

[patience111]

finally, find the solution on François Chollet's book (2nd edition), thank you.

[mattdangerw]

Hey! Thanks for opening this. Could you try pulling from the latest master? The easiest way would be something like pip install git+https://github.com/keras-team/keras-nlp.git --upgrade.

I believe we may have fixed your issue #270 and #292, in which case our next release will resolve this. Definitely worth checking.

Thanks!

[patience111]

Thank you for your suggestion, I'll try it later. Thank you!