diag_indices function issue?

https://github.com/numpy/numpy/blob/d9b1e32cb8ef90d6b4a47853241db2a28146a57d/numpy/lib/index_tricks.py#L883-L949

Hello there! Is it ok that diag_indices function returns two links on one index array?

For example, I’m trying to use this function to acces not only to the main diagonal but also to the diagonal with the given offset (like using offset parameter in np.diagonal). To implement that I do something like this:

X = np.array(...) # some 2D array
n, m = X.shape # without loss of generality assume that n <= m
xi, yi = np.diag_indices(n)
yi += offset
xi, yi = xi[np.logical_and(yi < m,yi >= 0)], yi[np.logical_and(yi < m, yi >= 0)]

print(xi, yi) # print to check if values are different (they aren't)
print(xi is yi) # print to check out if instances are different (they aren't)

To fix this issue I suggest to do the following:

...
return list(arange(n) for i in range(ndim))

[eric-wieser]

The benefit of having [np.arange(n)] * ndim wrapped up in it’s own function call

Note that (np.arange(n),) * ndim is different in an important way from what you say there, in that only the tuple version works for indexing (without a deprecation warning).

Agreed that this function is inconvenient though.

One thing we could do here is make the return value warn-on-write. Another would be to make the return value read-only.

I wonder if we can finally make np.diagonal return a read-write view, so that the need for this function vanishes.

[seberg]

Yeah, the diagonal thing has been warning since forever (at least it feels like it). I am happy with considering a full deprecation, or going the write-only/warn-on-write way. Going to the mailing list (best with a preferred/suggested choice maybe) sounds like a good idea.