Loaded Keras Model Throws Error While Predicting (Likely Issues with Masking)

Question

I am currently developing and testing a RNN that relies upon a large amount of data for training, and so have attempted to separate my training and testing files. I have one file where I create, train, and save a tensorflow.keras model to a file 'model.keras' I then load this model in another file and predict some values, but get the following error: Failed to convert elements of {'class_name': '__tensor__', 'config': {'dtype': 'float64', 'value': [0.0, 0.0, 0.0, 0.0]}} to Tensor. Consider casting elements to a supported type. See https://www.tensorflow.org/api_docs/python/tf/dtypes for supported TF dtypes

By the way, I have tried running model.predict with this exact same data in the file where I train the model, and it works smoothly. The model loading must be the problem, not the data used to predict.

This mysterious float64 tensor is the value I passed into the masking layer. I don't understand why keras is unable to recognize this JSON object as a Tensor and apply the masking operation as such. I have included snippets of my code below, edited for clarity and brevity:

model_generation.py:

# Create model

model = tf.keras.Sequential([
    tf.keras.layers.Input((352, 4)),
    tf.keras.layers.Masking(mask_value=tf.convert_to_tensor(np.array([0.0, 0.0, 0.0, 0.0]))),
    tf.keras.layers.GRU(50, return_sequences=True, activation='tanh'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.GRU(50,activation='tanh'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(units=1, activation='sigmoid')])

# Compile Model...
# Train Model...
model.save('model.keras')

model.predict(data) # Line works here

model_testing.py

model = tf.keras.models.load_model('model.keras')

model.predict(data) # this line generates the error

EDIT:

Moved the load command into the same file as the training, still receiving the exact same error message.

Answer 1

That error is due to the mask_value that you pass into tf.keras.layers.Masking not getting serialized compatibly for deserialization. But because you masking layer is a tensor containing all 0s anyway, you can instead just pass a scalar value like this and it will eliminate the need to serialize a tensor while storing the model

tf.keras.layers.Masking(mask_value=0.0)

and it broadcasts it to effectively make it equivalent to comparing it against the tensor containing all 0s. Here is the source where the mask is applied like this

ops.any(ops.not_equal(inputs, self.mask_value), axis=-1, keepdims=True)

and ops.not_equal supports broadcasting.