Keras: Simple math operation model fails with “Output tensors to a Model must be Keras tensors”

[hamzamerzic]

I want to define a simple model that uses basic math operations. I tried to implement it using multiple approaches, but the most obvious ones fail, and I would like to understand why.

Let's look at the code:

import keras.backend as K
from keras.layers import Input, Lambda
from keras.models import Model
import numpy as np

x = Input(shape=(3,))
y = Lambda(lambda x: x ** 2)(x)
print y
# Tensor("lambda_1/pow:0", shape=(?, 3), dtype=float32)
model = Model(inputs=x, outputs=y)
# Works!

y = x ** 2
print y
# Tensor("pow:0", shape=(?, 3), dtype=float32)
model = Model(inputs=x, outputs=y)
# Fails: TypeError: Output tensors to a Model must be Keras tensors.

y = K.pow(x, 2)
print y
# Tensor("Pow:0", shape=(?, 3), dtype=float32)
model = Model(inputs=x, outputs=y)
# Fails: TypeError: Output tensors to a Model must be Keras tensors. 

The last two examples fail due to:

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in __init__(self, inputs, outputs, name)
   1521             if not hasattr(x, '_keras_history'):
   1522                 cls_name = self.__class__.__name__
   1523                 raise TypeError('Output tensors to a ' + cls_name + ' must be '
-> 1524                                 'Keras tensors. Found: ' + str(x))
   1525         # Build self.output_layers:
   1526         for x in self.outputs:

As you can see, the output of y for all models is almost identical, but the intuitive ones fail for some reason. Any idea why this fails?

[farizrahman4u]

You can't use ops. You have to use layers. Which is why we have Lambda layer in the first place.

[fchollet]

Model(inputs=x, outputs=y)

Here both x and y should be the outputs of Keras layers. To use custom operations, you should wrap them in a Lambda layer (as @farizrahman4u points out), or write your own custom layer (see https://keras.io/layers/writing-your-own-keras-layers/)