Infinity loop when try to save model with input_signatures on funtion decorator about tensorflow HOT 3 OPEN

For simplify the bug reprodution follow the code where the inputs is two simple dict of tensors and one int tensor.
This code do the same as issue code and reproduce the error.

@tf.keras.saving.register_keras_serializable(package="MyLayers")

class BruteForce3(tf.keras.layers.Layer):

  def __init__(self, model, k=5, **kwargs):
      super().__init__(**kwargs)
      self.model = model
      self.k = k

  @tf.function(input_signature=[data_ds.element_spec, candidaster.element_spec, tf.TensorSpec([], dtype=tf.int32)])
  # @tf.function
  def call(self, queries, candidates, ks):
      
      standard_brute = tfrs.layers.factorized_top_k.BruteForce(self.model.query_model)
      
      candidates = [candidates]
      
      parse_one = pd.DataFrame.from_dict(candidates).to_dict(orient='list')
      candidates = tf.data.Dataset.from_tensor_slices(parse_one)
      
      standard_brute.index_from_dataset(tf.data.Dataset.zip(candidates.batch(2).map(lambda x: x['experience.jobTitle']), 
                                         candidates.batch(2).map(self.model.candidate_model)
                                        ))
      scores, ids = standard_brute(queries, ks)
      
      return scores, ids`

and save model:

`custom = BruteForce3(final_model)

for l in data_ds.take(1):
    scores, jobs = custom(l, cand[0], 2)
    
    print(scores, jobs)
    
    tf.saved_model.save(custom, "./serveModels/exportedMode36")`

when i save the model without function decorator and inputs_signatures i have the follow error on calling loaded model:
ValueError: Found zero restored functions for caller function.
and if i serve with tensorflow serving i have the follow response from server endpoint:
"error": "Serving signature name: \"serving_default\" not found in signature def"

from tensorflow.

Hi @Chm-vinicius ,

Could you please minimal code snippet for reproducing the issue? Thanks!

It seems the code contains tensorflow recommender layers which we are not supporting here at TF repo. I can see this is already reported in concerned repo.

If you can able to reproduce the issue with TF/Keras please submit the code snippet for same. Thanks!

from tensorflow.

Hi @SuryanarayanaY,

Thanks for your response!
Unfortunately due to NDA contract i cant share more than that, also we still in permantly contact with a google engineers to find and solve this issue.
When the solution are reached i certainly share her.

from tensorflow.