如何在 Keras 的输出中添加 tf.constant

Question

I have a running model, build with :

model = tf.keras.Model(inputs=input_layers, outputs=outputs)

If I try to add a simple constant to the outputs, I get an error message. ex:

output = output + [tf.constant(['label1', 'label2'], dtype = tf.string)]
model = tf.keras.Model(inputs=input_layers, outputs=outputs)

error message : AttributeError: Tensor.op is meaningless when eager execution is enabled.

Is there a way to add it to the model, even if after training or at the save() time.

The idea is to have the constant as an output during serving time.

example of a full network with the errror:

import tensorflow as tf
import tensorflow.keras as keras

input = keras.layers.Input(shape=(2,))
hidden = keras.layers.Dense(10)(input)
output = keras.layers.Dense(3, activation='sigmoid')(hidden)
model = keras.models.Model(inputs=input, outputs=[output, tf.constant(['out1','out2','out3'], dtype=tf.string)])

error

in <module>
      5 hidden = keras.layers.Dense(10)(input)
      6 output = keras.layers.Dense(3, activation='sigmoid')(input)
----> 7 model = keras.models.Model(inputs=input, outputs=[output, tf.constant(['out1','out2','out3'], dtype=tf.string)])

/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py in __init__(self, *args, **kwargs)
    144 
    145   def __init__(self, *args, **kwargs):
--> 146     super(Model, self).__init__(*args, **kwargs)
    147     _keras_api_gauge.get_cell('model').set(True)
    148     # initializing _distribution_strategy here since it is possible to call

/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/network.py in __init__(self, *args, **kwargs)
    165         'inputs' in kwargs and 'outputs' in kwargs):
    166       # Graph network
--> 167       self._init_graph_network(*args, **kwargs)
    168     else:
    169       # Subclassed network

/lib/python3.6/site-packages/tensorflow_core/python/training/tracking/base.py in _method_wrapper(self, *args, **kwargs)
    455     self._self_setattr_tracking = False  # pylint: disable=protected-access
    456     try:
--> 457       result = method(self, *args, **kwargs)
    458     finally:
    459       self._self_setattr_tracking = previous_value  # pylint: disable=protected-access

/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/network.py in _init_graph_network(self, inputs, outputs, name, **kwargs)
    268 
    269     if any(not hasattr(tensor, '_keras_history') for tensor in self.outputs):
--> 270       base_layer_utils.create_keras_history(self._nested_outputs)
    271 
    272     self._base_init(name=name, **kwargs)

/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/base_layer_utils.py in create_keras_history(tensors)
    182     keras_tensors: The Tensors found that came from a Keras Layer.
    183   """
--> 184   _, created_layers = _create_keras_history_helper(tensors, set(), [])
    185   return created_layers
    186 

/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/base_layer_utils.py in _create_keras_history_helper(tensors, processed_ops, created_layers)
    208     if getattr(tensor, '_keras_history', None) is not None:
    209       continue
--> 210     op = tensor.op  # The Op that created this Tensor.
    211     if op not in processed_ops:
    212       # Recursively set `_keras_history`.

/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py in op(self)
   1078   def op(self):
   1079     raise AttributeError(
-> 1080         "Tensor.op is meaningless when eager execution is enabled.")
   1081 
   1082   @property

AttributeError: Tensor.op is meaningless when eager execution is enabled.

using Python 3.6 and Tensorflow 2.0

Answer 1

Put the constant inside a Lambda layer. Keras does some extra book keeping so you need more than just tf operations for things to work. Using a Lambda layer will do this for you.

Edit to give an example of how this works: Your last example would translate to the following code

import tensorflow as tf
import tensorflow.keras as keras

inputs = keras.layers.Input(shape=(2,))
hidden = keras.layers.Dense(10)(inputs)
output1 = keras.layers.Dense(3, activation='sigmoid')(hidden)

@tf.function
def const(tensor):
    batch_size = tf.shape(tensor)[0]
    constant = tf.constant(['out1','out2','out3'], dtype=tf.string)
    constant = tf.expand_dims(constant, axis=0)
    return tf.broadcast_to(constant, shape=(batch_size, 3))

output2 = keras.layers.Lambda(const)(inputs)
model = keras.models.Model(inputs=inputs, outputs=[output1, output2])

Edit: This reminded me of a project I had a while back where I had to use a lot of constants in a Keras model. Back then I wrote a layer for it

class ConstantOnBatch(keras.layers.Layer):
    def __init__(self, constant, *args, **kwargs):
        self._initial_constant = copy.deepcopy(constant)
        self.constant = K.constant(constant)
        self.out_shape = self.constant.shape.as_list()
        self.constant = tf.reshape(self.constant, [1]+self.out_shape)
        super().__init__(*args, **kwargs)

    def build(self, input_shape):
        super().build(input_shape)

    def call(self, inputs):
        batch_size = tf.shape(inputs)[0]
        output_shape = [batch_size]+self.out_shape
        return tf.broadcast_to(self.constant, output_shape)

    def compute_output_shape(self, input_shape):
        input_shape = input_shape.as_list()
        return [input_shape[0]]+self.out_shape

    def get_config(self):
        base_config = super().get_config()
        base_config['constant'] = self._initial_constant

    @classmethod
    def from_config(cls, config):
        return cls(**config)

It might need some updating to tf2 and the code could definitely be written in a better way, but if you need a lot of constants this might provide a basis for a slightly more elegant solution than using a ton of Lambda layers.