Reshaping Numpy array repeating some elements

Question

I have trained a NN in Keras with LSTM, so I have been using 3D tensors. Now I want to predict with a dataset and I have to insert a 3D tensor in my NN.

(In my case I used features = 2 and lookback = 2, so input elements in LSTM are (batch_size, lookback, features))

So, imagine this example:

a = np.array([[1, 2],
              [3, 4]])

I need to do a_2 = np.reshape(1, 2, 2) to be able to insert it in the LSTM.

But if I have a bigger test dataset, like for instance:

b = np.array([[1, 2],
              [3, 4],
              [5, 6],
              [7, 8],
              [9, 10]])

I would need to convert it to a 3D array of this type:

b_2 = np.array([[[1, 2],
                 [3, 4]],

                [[3, 4],
                 [5, 6]],

                [[5, 6],
                 [7, 8]],

                [[7, 8],
                 [9, 10]]])

so in this case I have predictions for each lookback with the new point. I guess this could be done with a complicated solution using many for loops nested, but I wonder if there is more pythonic way. Thx.

Answer 1

Use numpy.stack() as follows.

>>> np.stack((b[:-1], b[1:]), axis=1)
array([[[ 1,  2],
        [ 3,  4]],

       [[ 3,  4],
        [ 5,  6]],

       [[ 5,  6],
        [ 7,  8]],

       [[ 7,  8],
        [ 9, 10]]])

Answer 2

You are looking for sliding windows and there's skimage's view_as_windows for that -

In [46]: from skimage.util.shape import view_as_windows

In [44]: features = 2; lookback = 2

In [45]: view_as_windows(b,(lookback, features))[:,0]
Out[45]: 
array([[[ 1,  2],
        [ 3,  4]],

       [[ 3,  4],
        [ 5,  6]],

       [[ 5,  6],
        [ 7,  8]],

       [[ 7,  8],
        [ 9, 10]]])