None is not supported in input_shape [bug]

[rex-yue-wu]

In the new release v2.0.1, None is not supported in input_shape, while None is must-have for a fully-convolutional network. Below is an example

>>> model = Sequential()
model = Sequential()
>>> model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(3, None, None)))
)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/nfs/isicvlnas01/share/anaconda/lib/python2.7/site-packages/keras/models.py", line 422, in add
    layer(x)
  File "/nfs/isicvlnas01/share/anaconda/lib/python2.7/site-packages/keras/engine/topology.py", line 528, in __call__
    self.build(input_shapes[0])
  File "/nfs/isicvlnas01/share/anaconda/lib/python2.7/site-packages/keras/layers/convolutional.py", line 125, in build
    raise ValueError('The channel dimension of the inputs '
ValueError: The channel dimension of the inputs should be defined. Found `None`.
>>> print keras.__version__
print keras.__version__
2.0.1

[kivantium]

Your problem is the order of input_shape. You assume channels come first, but whether channels come first or last depends on the configuration in $HOME/.keras/keras.json. See https://keras.io/backend/.

If your image_data_format is channels_last, your input_shape should be (None, None, 3).

[rex-yue-wu]

Nope. I use theano backend, and the order should be ( 3, None, None ).

[bstriner]

Tried your code and works for me on Theano. Double check the ordering actually loaded in keras.

import keras.backend as K
print(K.image_dim_ordering())

'th'

Cheers

[bstriner]

Just realized I gave you the old API.

import keras.backend as K
print(K.image_dim_ordering())
print(K.image_data_format())

th
channels_first

[rex-yue-wu]

Thank you. As you both said, the image_data_format caused my error. The real problem is that I did not update my keras.json for the new version, because image_data_format is not in a 1.x version.