I need to augment(replicate) a 2d array of shape 32X32 to a 3d array of shape 32X32X3 by duplicating the source array. How can i do this in the best possible way?
Below is the sample of the source and expected array. I need to apply this logic over a bigger scope of my application
Source array:
array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) Expected array:
array([[[1, 2, 3], [4, 5, 6], [7, 8, 9]], [[1, 2, 3], [4, 5, 6], [7, 8, 9]], [[1, 2, 3], [4, 5, 6], [7, 8, 9]]]) 3 Answers
By my tests, np.repeat is a little faster than np.tile:
X = np.repeat(arr[None,:], 3, axis=0) Alternatively, use np.concatenate:
X = np.concatenate([[arr]] * 3, axis=0) arr = np.arange(10000 * 1000).reshape(10000, 1000) %timeit np.repeat(arr[None,:], 3, axis=0) %timeit np.tile(arr, (3, 1, 1)) %timeit np.concatenate([[arr]] * 3, axis=0) # Read-only, array cannot be modified. %timeit np.broadcast_to(arr, (3, *arr.shape)) # Creating copy of the above. %timeit np.broadcast_to(arr, (3, *arr.shape)).copy() 170 ms ± 3.82 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) 187 ms ± 3.12 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 243 ms ± 3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) 10.9 µs ± 218 ns per loop (mean ± std. dev. of 7 runs, 100000 loops 189 ms ± 2.45 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)each) np.array_equals(np.repeat(arr[None,:], 3, axis=0), np.tile(arr, (3, 1, 1)) True Sounds like a job for np.tile:
In [101]: np.tile(A, (3,1,1)) Out[101]: array([[[1, 2, 3], [4, 5, 6], [7, 8, 9]], [[1, 2, 3], [4, 5, 6], [7, 8, 9]], [[1, 2, 3], [4, 5, 6], [7, 8, 9]]]) The second argument specifies the number of copies on each dimension.
0If you don't need to modify the result, make use of broadcast_to:
np.broadcast_to(arr, (3, *arr.shape)) Validation using @coldspeed's answer:
arr = np.arange(10000 * 1000).reshape(10000, 1000) X = np.repeat(arr[None,:], 3, axis=0) broadcast_x = np.broadcast_to(arr, (3, *arr.shape)) np.array_equal(X, broadcast_x) True If you do need to be able to modify, you can call copy() on the result, which should come close to repeat and tile in terms of speed.
