Dimension Mismatch in LSTM Keras

Question

I want to create a basic RNN that can add two bytes. Here are the input and outputs, which are expected of a simple addition

X = [[0, 0], [0, 1], [1, 1], [0, 1], [1, 0], [1, 0], [1, 1], [1, 0]]

That is, X1 = 00101111 and X2 = 01110010

Y = [1, 0, 1, 0, 0, 0, 0, 1]

I created the following sequential model

model = Sequential()
model.add(GRU(output_dim = 16, input_length = 2, input_dim = 8))
model.add(Activation('relu'`))
model.add(Dense(2, activation='softmax'))
model.compile(loss = 'binary_crossentropy', optimizer = 'adam', metrics = ['accuracy'])
model.summary()

The error I get is something along

expected lstm_input_1 to have 3 dimensions, but got array with shape (8L, 2L)

So if I increase the dimensions by changing X to

[[[0 0]] [[1 1]] [[1 1]] [[1 0]] [[0 0]] [[1 0]] [[0 1]] [[1 0]]]

Then the error changes to

expected lstm_input_1 to have shape (None, 8, 2) but got array with shape (8L, 1L, 2L)

Answer 1

In Keras the Sequential models expect an input of shape (batch_size, sequence_length, input_dimension) . I suspect you need to change the two last dimensions of your input array. Remember, the batch dimension is not explicitly defined.

Answer 2

将X更改为[[[0, 0], [0, 1], [1, 1], [0, 1], [1, 0], [1, 0], [1, 1], [1, 0]]]使其形状为(1, 8, 2)

Answer 3

Keras as input requiers 3D data, as stated in error. It is samples, time steps, features. Since you have (8L, 2L) Keras takes it as 2D - [samples, features]. In order to fix it, do something like this

def reshape_dataset(train):
    trainX = numpy.reshape(train, (train.shape[0], 1, train.shape[1]))
    return numpy.array(trainX)

x = reshape_dataset(your_dataset)

now X should be 8L,1,2L which is [samples, time steps, features] - 3D