ConvLSTM2D: return_sequences=False returns a 5D tensor (should return 4D)

According to ConvLSTM2D documentation, when return_sequences=False, the ConvLSTM2D is supposed to only return the last element of the sequence. Specifically:

     # Output shape
        - if `return_sequences`
             - if data_format='channels_first'
                5D tensor with shape:
                `(samples, time, filters, output_row, output_col)`
             - if data_format='channels_last'
                5D tensor with shape:
                `(samples, time, output_row, output_col, filters)`
        - else
            - if data_format ='channels_first'
                4D tensor with shape:
                `(samples, filters, output_row, output_col)`
            - if data_format='channels_last'
                4D tensor with shape:
                `(samples, output_row, output_col, filters)`
            where o_row and o_col depend on the shape of the filter and
            the padding

But setting return_sequences=False seem to have no effect and the layer continues to return a 5-d tensor.

Is this a bug? Is there a workaround?

Here is my code:

def create_model(input_shape):

    num_filters = 15
    model = Sequential()
    model.add(ConvLSTM2D(filters=num_filters, kernel_size=(3, 3),
                         # input_shape=(None, 40, 40, 1),
                         input_shape=input_shape,
                         padding='same', return_sequences=True))
    model.add(BatchNormalization())

    model.add(ConvLSTM2D(filters=num_filters/2, kernel_size=(3, 3),
                         padding='same', return_sequences=True))
    model.add(BatchNormalization())

    model.add(ConvLSTM2D(filters=num_filters, kernel_size=(3, 3),
                         padding='same', return_sequences=False))
    model.add(BatchNormalization())

    model.add(Conv2D(filters=1, kernel_size=(3, 3),
                     activation='sigmoid',
                     padding='same', data_format='channels_last'))

    opt = optimizers.Adam(lr=1e-4)
    model.compile(loss='mae', optimizer=opt, metrics=[metrics.mae])

    return model

[sunmonk]

hey. seq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3), padding=‘same’, data_format=‘channels_last’, return_sequences=False))

filters should not be 40 but 3!

seq.add(ConvLSTM2D(filters=3, kernel_size=(3, 3), padding=‘same’, data_format=‘channels_last’, return_sequences=False))

[fchollet]

Can’t reproduce, in this script, the output of the last ConvLSTM2D is a 4D tensor with shape (None, 40, 40, 40).

_standardize_input_data

It sounds like the problem is with the data that you are passing as input/target.