Flatten functional API -- can't pass output to Dense -- shape calculation issue?

[David S]

I am trying to use the keras functional API to flatten a  batch of images and apply a dense layer. However when I do

flatten()(img_batch)

I get something that doesn't know its shape, that is, I get

In [34]: keras.layers.Flatten()(yy)
Out[34]: <tf.Tensor 'Reshape:0' shape=(?, ?) dtype=float32>

Above, yy is a batch of images that are (76,56,1), I'd hope that the shape from Flatten() would be (?, 4256)

I think this is why the subsequent call to the Dense layer behaves as follows:

 

File "xtcav_new.py", line 182, in autoencoder
    encoded = Dense(encoding_dim, activation='relu')(flat)
  File "/home/davidsch/miniconda2/envs/mlearn/lib/python2.7/site-packages/keras/engine/topology.py", line 543, in __call__
    self.build(input_shapes[0])
  File "/home/davidsch/miniconda2/envs/mlearn/lib/python2.7/site-packages/keras/layers/core.py", line 752, in build
    constraint=self.W_constraint)
  File "/home/davidsch/miniconda2/envs/mlearn/lib/python2.7/site-packages/keras/engine/topology.py", line 415, in add_weight
    weight = initializer(shape, name=name)
  File "/home/davidsch/miniconda2/envs/mlearn/lib/python2.7/site-packages/keras/initializations.py", line 59, in glorot_uniform
    s = np.sqrt(6. / (fan_in + fan_out))
TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'

I was trying to modify the keras autoencoder blog at https://blog.keras.io/building-autoencoders-in-keras.html for a dataset I have. This is keras 1.2.0 and Tensorflow 0.12.

best,

David

[Benjamin Striner]

So I assume yy is (None, 76,56,1), such as gray 76x56 images. The below works for me:

from keras.layers import Dense, Input, Flatten

from keras.models import Model

a = Input((76,56,1))

b = Flatten()(a)

c = Dense(100)(b)

m = Model(a,c)

m.summary()

____________________________________________________________________________________________________

Layer (type)                     Output Shape          Param #     Connected to

====================================================================================================

input_1 (InputLayer)             (None, 76, 56, 1)     0

____________________________________________________________________________________________________

flatten_1 (Flatten)              (None, 4256)          0           input_1[0][0]

____________________________________________________________________________________________________

dense_1 (Dense)                  (None, 100)           425700      flatten_1[0][0]

====================================================================================================

Total params: 425,700

Trainable params: 425,700

Non-trainable params: 0