Allow to reshape unknown tensors.

[hgaiser]

The current implementation of keras.layers.Reshape does not allow the reshaping of tensors of unknown shape (ie. (None, None, 3) for any arbitrary image). Strangely enough, the underlying code of tensorflow.reshape does allow this kind of reshape (and already handles -1 in the target shape), I'm not sure about other backends.

Motivation
The following code shows what this PR fixes:

import keras

inputs = keras.layers.Input((None, None, 3))
outputs = keras.layers.Reshape((-1, 1))(inputs)

model = keras.models.Model(inputs=inputs, outputs=outputs)

print("Success!")

Without this PR the result is:

Using TensorFlow backend.
Traceback (most recent call last):
  File "test.py", line 4, in <module>
    outputs = keras.layers.Reshape((-1, 1))(inputs)
  File "/home/hgaiser/.local/lib/python3.6/site-packages/Keras-2.0.8-py3.6.egg/keras/engine/topology.py", line 602, in __call__
  File "/home/hgaiser/.local/lib/python3.6/site-packages/Keras-2.0.8-py3.6.egg/keras/layers/core.py", line 391, in call
  File "/home/hgaiser/.local/lib/python3.6/site-packages/Keras-2.0.8-py3.6.egg/keras/layers/core.py", line 376, in compute_output_shape
  File "/home/hgaiser/.local/lib/python3.6/site-packages/Keras-2.0.8-py3.6.egg/keras/layers/core.py", line 364, in _fix_unknown_dimension
  File "/home/hgaiser/.local/lib/python3.6/site-packages/numpy/core/fromnumeric.py", line 2518, in prod
    out=out, **kwargs)
  File "/home/hgaiser/.local/lib/python3.6/site-packages/numpy/core/_methods.py", line 35, in _prod
    return umr_prod(a, axis, dtype, out, keepdims)
TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'

With this PR the result is:

Using TensorFlow backend.
Success!

ps. my current workaround is:

outputs = keras.layers.Lambda(lambda x: keras.backend.reshape(x,
 (keras.backend.shape(x)[0],) + (-1, 1)))(inputs)

[fchollet]

I believe the current design was due to previous limitations with TF.

Tests are currently failing. Also, you should add a unit test demonstrating the limitation that is being lifted.

[fchollet]

Please add a comment describing what this does / why do it, and make this line shorter.

[hgaiser]

Done.

[fchollet]

Since this function is only called once, consider in-lining it (not saying you should do it, but consider whether it would improve the code if you did it -- maybe it would).

[hgaiser]

I had considered that, but I figured keeping it separate would be better. That way it can potentially be used by layers that subclass Reshape (or other code in general). In addition it is more in line with the "separation of concerns" philosophy.

[hgaiser]

Updated PR as requested.

[fchollet]

Tests are failing: https://travis-ci.org/fchollet/keras/builds/279680364?utm_source=github_status&utm_medium=notification

[fchollet]

LGTM, thanks