I am trying to create an numpy array(lets say A) of shape (a, b, 1, d) where d is unknown and changes according to the input I have. I have another array (lets say B) with shape (a, b, 1, 1). I want to append the values of B to A from the for loop. In matlab it can be easily done with:
a = zeros(a, b, 1, 1)
count = 0
for i = 1 : something
ai = array of shape (a, b, 1, 1)
count += 1
a(:, :, 1, count) = ai
end
How can I achieve similar result in python ?
MATLAB can grow matrices simply by indexing new values; numpy does not allow that. But you can concatenate A and B creating a new array with shape (a,b,1,d+1)
In [1187]: np.concatenate((np.ones((2,3,1,4)), np.ones((2,3,1,1))), axis=-1).shape
Out[1187]: (2, 3, 1, 5)
But if you want to do this multiple times, I'd suggest collecting the intermediate arrays in a list, and do one concatenate at the end. That's more efficient.
In [1189]: a = [np.zeros((2,3,1,1))]
In [1190]: for i in range(4):
...: a.append(np.ones((2,3,1,1))*i)
...: a = np.concatenate(a, axis=-1)
...:
In [1191]: a.shape
Out[1191]: (2, 3, 1, 5)
In [1192]: a
Out[1192]:
array([[[[ 0., 0., 1., 2., 3.]],
....
[[ 0., 0., 1., 2., 3.]]]])
numpy.append can be used, but it is not efficient as when used in a loop you get a code with quadratic time complexity. So the problem is "how to do it efficiently".
It seems that the standard overallocation technique to gain O(1) amortized complexity of the append operation is used in the dynarray package with is just light weight wrapper to the np.array.
somethingknown at runtime?