[Numpy-discussion] Re: Add diagonal offset argument to all functions that are missing it
Ralf Gommers wrote: > This sounds quite reasonable to me. The `k=0` keyword is quite badly named, > which is my one concern. Especially when tacking it on at the end of a > signature with already 3-4 keywords, it's not a good name. How about > something like `diag_offset`? FWIW, we chose `offset` for `array_api_extra.create_diagonal`, instead of inheriting `k` from `np.diag`. `np.diagonal` and `np.linalg.trace` also use `offset`. Given that all of the proposed functions apart from `np.identity` already have "diag" as a substring of their name, I think just `offset` would be fine. What else could `offset` mean in the case of `np.identity`? I suppose there is an argument for actually leaving `np.identity` as is—if someone wants a square off-diagonal matrix of ones, which isn't an identity matrix, their code might be more readable with `np.diag(np.ones(...), k=1)` or just `np.eye` instead, right? Maybe it is weird for "eye" but not "identity" to have this capability, though. Cheers, Lucas ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Re: Making `T` property Array API compatible
> The new discrepancy between `arr.T` and `arr.transpose()` is justified, as > `T` is defined by the Array API, where `transpose` isn't and should retain > the existing behavior. The other side of the coin here is that this change would fix the discrepancy between `arr.T` and the functions `np.matrix_transpose` and `np.linalg.matrix_transpose`, which implement batched transpose over matrices in the 2 innermost dimensions, rather than reversing all axes. In [*10*]: X = np.stack((np.eye(2), np.eye(2))) In [*12*]: X.T Out[*12*]: array([[[1., 1.], [0., 0.]], [[0., 0.], [1., 1.]]]) In [*13*]: np.matrix_transpose(X) Out[*13*]: array([[[1., 0.], [0., 1.]], [[1., 0.], [0., 1.]]]) Cheers, Lucas On 12 Apr 2025, at 11:14, Mateusz Sokol wrote: Hi all! The Array API standard states that `T` property should only be applied to 2-dimensional arrays, in all other cases it should raise an error: https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.T To ensure that NumPy also follows this rule, I opened a PR that raises a warning for `arr.T` for non-2-dimensional arrays and scalars: https://github.com/numpy/numpy/pull/28678 For non-2-dimensional arrays, the replacement for `arr.T` can be either: Array API compatible, namely `np.permute_dims(arr, range(arr.ndim)[::-1])`, or shorter, NumPy specific: `arr.transpose()`. The new discrepancy between `arr.T` and `arr.transpose()` is justified, as `T` is defined by the Array API, where `transpose` isn't and should retain the existing behavior. Please share your thoughts! Best regards, Mateusz ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: lucas.coll...@gmail.com ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
[Numpy-discussion] Re: Making `T` property Array API compatible
If NumPy were to make a move on the deprecation, then I think it would be reasonable to change the standard from > If the array instance is not two-dimensional, an error should be raised. to “if the array instance is not two-dimensional, behaviour should match `.mT`, or an error should be raised.” > On 14 Apr 2025, at 08:35, Sebastian Berg wrote: > > On Sat, 2025-04-12 at 10:10 +, Mateusz Sokol wrote: >> Hi all! >> >> The Array API standard states that `T` property should only be >> applied to 2-dimensional arrays, in all other cases it should raise >> an error: >> https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.T >> >> To ensure that NumPy also follows this rule, I opened a PR that >> raises a warning for `arr.T` for non-2-dimensional arrays and >> scalars: https://github.com/numpy/numpy/pull/28678 > > > There was once a surprising amount of resistance to doing this exact > change a long time ago in NumPy and that is exactly why it never > happened earlier and that is why we have `.mT` and not just `.T` to > begin with. > > I am still happy with slowly deprecating it with a message to use > `arr.transpose()`, `np.moveaxis()`, or `.mT` when it is applies. > Maybe making sure that `.T` keeps working at least for 2-D, and think > about what to do for 1-D (probably an error), although I am not sure > about this unless there is a long term plan to consider allowing `.T` > to mean the same as `.mT`. > > > That sais, it is not correct to say there is any incompatibility! This > is an opinionated recommendation at most and if it was more it would > probably be a mistake there. > > - Sebastian > > > >> >> For non-2-dimensional arrays, the replacement for `arr.T` can be >> either: Array API compatible, namely `np.permute_dims(arr, >> range(arr.ndim)[::-1])`, or shorter, NumPy specific: >> `arr.transpose()`. >> >> The new discrepancy between `arr.T` and `arr.transpose()` is >> justified, as `T` is defined by the Array API, where `transpose` >> isn't and should retain the existing behavior. >> >> Please share your thoughts! >> >> Best regards, >> Mateusz >> ___ >> NumPy-Discussion mailing list -- numpy-discussion@python.org >> To unsubscribe send an email to numpy-discussion-le...@python.org >> https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ >> Member address: sebast...@sipsolutions.net >> > > ___ > NumPy-Discussion mailing list -- numpy-discussion@python.org > To unsubscribe send an email to numpy-discussion-le...@python.org > https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ > Member address: lucas.coll...@gmail.com ___ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
