# Numpy array slicing/reshape/concatination

I'm sure my question exists on the internet, i just don't know the right formulations.

I have a data-sample as input for a NN. This sample has the shape (1, 8, 28, 80). Basically it is 80 timesteps of an image. -> y=8, x=28, time=80

i can extract the image at time=0 with:

np_img = image.data.numpy()   # shape (1, 8, 28, 80)
t0 = np_img[:, :, :, 0]


in order to be able to plot the images at each timestamp below each other, resulting in an array of (640, 28), ergo concatenating along the y-axis I do:

amount_timeslots = img.shape[-1]
new_array = img[:, :, :, 0]

for i in range(1, amount_timeslots):
ti = img[:, :, :, i]
new_array = np.concatenate((new_array, ti))

new_array.shape  # (640, 28)


Is there a more pythonic way by using build in numpy magic to do this?

• @MaartenFabré that returns (28,640) and the values seem to be not in the right order (so .T does help changing the dimensions but the values are wrong) – Maikefer Jun 3 '20 at 13:30

# concatenate

There is no need to do the concatenation pairwise np.concatenate([img[:, :, :, i] for i in range(img.shape[-1])]) should improve the speed already

# numpy.moveaxis

You can use numpy.moveaxis

new_array2 = np.concatenate(np.moveaxis(img,(0,1,2), (1,2,0)))


To check whether the result is the same:

assert np.array_equal(new_array, new_array2)


# Other improvements

Since you don't use the first axis of the image, you can do np_img = image.data.numpy() so prevent you repeating the  in all subsequent code