Saving a composite model that includes a custom layer results in error - None has NoneType, but expected one of: bytes, unicode

Question

I'm trying to save a model which is a composite model of composite models.

The first model is a sequential model of two sequential models. Both of the two sub-models have custom layers that perform scaling operations like MinMax and cube root. This model saves and loads without any issues.

The first model is then loaded in a different script without compiling. There is not issue with this. Let's call this model MODEL_1.

The next step may be a bit confusing. There are then two more models added in parallel to each other but sequentially with MODEL_1. Let's call these models MODEL_2a and MODEL_2b. The output of MODEL_1 has the input of MODEL_1 concatenated to it and this serves as the input to MODEL_2a and MODEL_2b. It should be noted that that the "MODEL_1 input" goes through a custom scaling layer before being concatenated to the output of MODEL_1. The is scaling layer that was also implemented in MODEL_1 without any issues.

Finally, there is a single, custom layer that performs a 'simple' weighted sum of the outputs of MODEL_2a and MODEL_2b to produce the model output. This weighting happens via alpha*OUTPUT_2a + [1-alpha]*OUTPUT_2b. 'alpha' is a trainable, scalar parameter. I haven't used this before in a model that I have saved, so I'm guessing this is the cause.

The model compiles and trains, but fails to save.

The custom weighted sum layer is this,

class WeightedSum(krs.layers.Layer):
    def __init__( self, n_models = 2, name = 'weighted_sum_0' ):
        super( WeightedSum, self ).__init__( name = name)
        self.n_models = n_models
        self.ensemble_weights = []
        self.output_init = tf.Variable(0.,validate_shape=False,trainable=False)

    def build(self,input_shape):
        for i in range(self.n_models):
            self.ensemble_weights.append( self.add_weight(shape=(1,),
                                    initializer = 'ones',
                                    trainable = True) )

    def call(self,inputs):
        new_normalizer = tf.convert_to_tensor(0.,dtype = inputs[0].dtype)
        for i in range(self.n_models):
            new_normalizer = new_normalizer + self.ensemble_weights[i]
        new_normalizer = tf.constant(1.,dtype=new_normalizer.dtype)/new_normalizer
        output = self.output_init

        for i in range(self.n_models):
            output = tf.add(output,tf.multiply(self.ensemble_weights[i],inputs[i]))
        output = tf.multiply( output, new_normalizer )
        return output

The save command is this, (NOTE: I use import tensorflow.keras as krs)

krs.models.save_model(linked_model,"test_failed_save.mdl")

The error that is produced is this.

Traceback (most recent call last):
  File "multi_fidelity_training_full_link.py", line 304, in <module>
    main()
  File "multi_fidelity_training_full_link.py", line 265, in main
    krs.models.save_model(linked_model,"test_failed_save.mdl")
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/keras/saving/save.py", line 151, in save_model
    signatures, options, save_traces)
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/keras/saving/saved_model/save.py", line 90, in save
    model, filepath, signatures, options)
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/saved_model/save.py", line 1104, in save_and_return_nodes
    raise_metadata_warning))
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/saved_model/save.py", line 1291, in _build_meta_graph
    raise_metadata_warning)
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/saved_model/save.py", line 1225, in _build_meta_graph_impl
    options.namespace_whitelist)
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/saved_model/save.py", line 713, in _fill_meta_graph_def
    _call_function_with_mapped_captures, resource_map=resource_map)))
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/training/tracking/graph_view.py", line 424, in frozen_saveable_objects
    call_with_mapped_captures)
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/training/tracking/graph_view.py", line 375, in _serialize_gathered_objects
    slot_variables=slot_variables)
  File "/usr/WS2/mvander/py3venv/py3venv/lib/python3.7/site-packages/tensorflow/python/training/tracking/graph_view.py", line 355, in _fill_object_graph_proto
    child_proto.local_name = child.name
TypeError: None has type NoneType, but expected one of: bytes, unicode

How do I fix this?

Answer 1

when you use add_weight, give a name.[self.add_weight('myname', shape=....)]