Concatenate input with constant vector in keras

Question

I am trying to concatenate my input with a constant tensor in the keras-2 function API. In my real problem, the constants depend on some parameters in setup, but I think the example below shows the error I get.

from keras.layers import*
from keras.models import *
from keras import backend as K
import numpy as np

a = Input(shape=(10, 5))
a1 = Input(tensor=K.variable(np.ones((10, 5))))
x = [a, a1]  # x = [a, a] works fine
b = concatenate(x, 1)
x += [b]  # This changes b._keras_history[0].input
b = concatenate(x, 1)
model = Model(a, b)

The error I get is:

ValueError                                Traceback (most recent call last)
~/miniconda3/envs/ds_tools/lib/python3.6/site-packages/keras/engine/topology.py in assert_input_compatibility(self, inputs)
    418             try:
--> 419                 K.is_keras_tensor(x)
    420             except ValueError:

~/miniconda3/envs/ds_tools/lib/python3.6/site-packages/keras/backend/theano_backend.py in is_keras_tensor(x)
    198                           T.sharedvar.TensorSharedVariable)):
--> 199         raise ValueError('Unexpectedly found an instance of type `' + str(type(x)) + '`. '
    200                          'Expected a symbolic tensor instance.')

ValueError: Unexpectedly found an instance of type `<class 'theano.gpuarray.type.GpuArraySharedVariable'>`. Expected a symbolic tensor instance.

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
<ipython-input-2-53314338ab8e> in <module>()
      5 a1 = Input(tensor=K.variable(np.ones((10, 5))))
      6 x = [a, a1]
----> 7 b = concatenate(x, 1)
      8 x += [b]  # This changes b._keras_history[0].input
      9 b = concatenate(x, 1)

~/miniconda3/envs/ds_tools/lib/python3.6/site-packages/keras/layers/merge.py in concatenate(inputs, axis, **kwargs)
    506         A tensor, the concatenation of the inputs alongside axis `axis`.
    507     """
--> 508     return Concatenate(axis=axis, **kwargs)(inputs)
    509 
    510 

~/miniconda3/envs/ds_tools/lib/python3.6/site-packages/keras/engine/topology.py in __call__(self, inputs, **kwargs)
    550                 # Raise exceptions in case the input is not compatible
    551                 # with the input_spec specified in the layer constructor.
--> 552                 self.assert_input_compatibility(inputs)
    553 
    554                 # Collect input shapes to build layer.

~/miniconda3/envs/ds_tools/lib/python3.6/site-packages/keras/engine/topology.py in assert_input_compatibility(self, inputs)
    423                                  'Received type: ' +
    424                                  str(type(x)) + '. Full input: ' +
--> 425                                  str(inputs) + '. All inputs to the layer '
    426                                  'should be tensors.')
    427 

ValueError: Layer concatenate_2 was called with an input that isn't a symbolic tensor. Received type: <class 'theano.gpuarray.type.GpuArraySharedVariable'>. Full input: [concatenate_1/input_3, concatenate_1/variable]. All inputs to the layer should be tensors.

I am running keras version 2.0.5 with the theano backend, with theano version 0.10.0dev1. Any ideas on what is going wrong or a more correct way to accomplish the concatenation?

Answer 1

Dimensions in keras work like this:

• When you define them in layers, building your model, you never define "batch_size".
• Internally, using backend functions, in loss functions and in any tensor operation, the batch dimension is the first

Keras shows you a None to represent the batch size in summaries, errors and others.

That means that:

• a's shape is (None, 10, 5)
• a1's shape just (10,5). You cannot concatenate them.

There are a few workarounds you can do, such as creating a1 with shape (1,10,5) and then repeating it's values in the batch dimension:

constant=K.variable(np.ones((1,10, 5)))
constant = K.repeat_elements(constant,rep=batch_size,axis=0)

I was totally unable to use Input(tensor=...) because the constant's dimension is fixed, and the input's dimension is None, so I worked it around with a lambda layer:

b = Lambda(lambda x: K.concatenate([x,constant],axis=1),output_shape=(20,5))(a)

But I can't at all understand what you want to achieve with x += [b] and the rest.