Can't get flatten() to work with tensorflow backend

[Isaac Gerg]

I have a simple example which works in Theano but breaks tensorflow.

input = keras.layers.Input(batch_shape=(batch_size,400,400, 1))

x = AveragePooling2D(pool_size=(2,2))(input)

# <tf.Output 'AvgPool:0' shape=(256, 200, 200, 1) dtype=float32>

x = Flatten()(x)

# <tf.Output 'Reshape:0' shape=(?, ?) dtype=float32>

x has unknown shape to tensorflow.  This causes issues further down the line.

What am I doing wrong here?

[yohanes...@gmail.com]

One of the common mistake related to Flatten() is related to keras image_dim_ordering configuration https://keras.io/backend/. Have you configured properly (tf for tensorflow and th for theano) ?

If that is not your problem, you might want to post the error message as well so other can understand the problem better.

[Isaac Gerg]

Yes, I did check this as I initially thought htis was the issue.  The ticket is here: https://github.com/fchollet/keras/issues/4470

Here is the code that reproduces the problem (this code only breaks in tensorflow).

    input = keras.layers.Input(shape=(400,400, 1))

    x = AveragePooling2D(pool_size=(2,2))(input)

    x = Flatten()(x)  # This flatten work fine.

    x = keras.layers.RepeatVector(3)(x)

    x = Reshape((200, 200, 3))(x)

    y = AveragePooling2D(pool_size=(4,4))(input)

    y = Flatten()(y)  # This flatten works fine.

    y = keras.layers.RepeatVector(3)(y)

    y = Reshape((100, 100, 3))(y)

    vgg1 = keras.applications.vgg16.VGG16(include_top=False)

    vgg1.trainable = False    # Doesnt work

    x = vgg1(x)

    vgg2 = keras.applications.vgg16.VGG16(include_top=False)

    vgg2.trainable = False  # Doesnt work

    y = vgg2(y)

    yUp  = keras.layers.UpSampling2D((2,2))(y)

    m = keras.layers.merge([x,yUp], mode='sum')

    m = keras.layers.Convolution2D(64,3,3, border_mode='same')(m)

    m = keras.layers.Flatten()(m)  # This line is the error, replacing it with the next line fixes it.

    #m = keras.layers.Reshape((6*6*64,))(m)

    m = keras.layers.MaxoutDense(8)(m)

    m = keras.layers.Dropout(0.5)(m)

    m = keras.layers.MaxoutDense(4)(m)

    m = keras.layers.Dropout(0.5)(m)

    m = keras.layers.Dense(2, activation='softmax')(m)

    model = keras.models.Model(input, m)

    return model