[FEATURE REQUEST] `axes` argument for `ndarray.transpose()`

[sfe-SparkFro]

Describe the solution you'd like

In standard NumPy, ndarray.transpose() accepts an optional axes argument. For example

import numpy as np

my_array = np.zeros((1,2,3))
transposed_array = my_array.transpose((2, 0, 1))

print("Original array shape:", my_array.shape)
print("Transposed array shape:", transposed_array.shape)

Prints:

Original array shape: (1, 2, 3)
Transposed array shape: (3, 1, 2)

[v923z]

Is this feature needed? Obviously, we can't re-implement everything that's in numpy, so we've got to be selective, and work on functions that are actually necessary in an application.

[sfe-SparkFro]

I'm using an array to represent an image, so it needs 3 dimensions (rows, columns, channels) in that order. I need to rotate the image in increments of 90 degrees. The fastest way to do that is to transpose the first 2 axes (columns, rows, channels) then flip one of those first 2 axes (just a change of the strides and shape).

Without being able to specify the axes, transposing results in (channels, columns, rows), which is not what's needed. I'm open to alternative solutions, as long as performance is good with minimal code (eg. can't copy the channels one by one).

[v923z]

Can you post a python snippet that would do what you want? I don't mean numpy code, I mean raw python. Is this something that can be achieved by manipulating the strides, or are there cases, when you actually have to copy data?

[sfe-SparkFro]

Can you post a python snippet that would do what you want? I don't mean numpy code, I mean raw python.

I'll have to get back to you, I'm traveling right now.

Is this something that can be achieved by manipulating the strides, or are there cases, when you actually have to copy data?

I think only the strides and shape need to change, no copying data should be required to transpose axes.

Is it possible to change the strides directly in MicroPython? Can you do something like my_array.strides = (2,3,1)?