I am familiar with that issue and many older ones in this mailing list too. The argument I am trying to make is that just because it is a view should not directly imply that it should go in the NumPy main namespace. I don't know what array API designers think but .H is order of magnitude more common than tensor transpose. numpy.sort is also inplace and also tricky but we have it. Plus ".H" can have the correct transpose for the tensors and can behave like matrix_transpose for reals or whatever. The point I was trying to make is that this array API spec should also involve usability aspects of the tool and if complex operations are not discussed in array API then either this API spec is incomplete or it is a float array API. But like I said, I don't want to start any discussion, did a bit too much spectral work on time-series lately so finger-wounds are still fresh I guess. Apologies for the rant. On Thu, Dec 8, 2022 at 11:54 PM Aaron Meurer <asmeurer@gmail.com> wrote:
On Wed, Dec 7, 2022 at 3:49 PM Ilhan Polat <ilhanpolat@gmail.com> wrote:
On matrix_transpose() : Every time this discussion brought up, there was a huge resistance to
add more methods to array object or new functions (I have been involved in some of them on the pro .H side, links you have given and more in the mailing list) and now we are adding .mT and not .H? That is very surprising to me (and disappointing) after forcing people to write A.conj().T for years which is fundamentally the most common 2D operation regarding transpose and from a user experience perspective. Having a function name with the word "matrix" is already problematic but I can live with that. But adding .mT to the main namespace seems really going against all decisions made in the past. I also wish for cache-oblivious inplace transposition too that would make many linalg functions perform faster but I wouldn't dare to propose inplace_transpose() or .iT because it is not that important for *all* users. And in a way, neither is mT.
Again not trying to starting an old dumpster fire but surely there must
have been some consideration for .H before we ended up with milliTranspose. In fact, as I am typing this, I am already regretting it. Just a rant about typing too much conj().T lately I guess.
.H was discussed in this issue https://github.com/numpy/numpy/issues/13797
The problem with .H is that it wouldn't be a view, since it takes a conjugate. Some ideas were suggested to fix this, but they are much more nontrivial to implement, and it's not even clear if they are desired (basically you'd need a new complex conjugate dtype). x.mT on the other hand can easily be a view, since it's basically just a shorthand for swapaxes(x, -1, -2).
More to the point, my plan here is only to work on functions that are part of the array API specification (and possibly extending these features to related things like adding namedtuples for other functions if there are any). Hermitian transpose has not yet been discussed for addition to the array API specification, which only recently gained support for complex numbers.
Aaron Meurer
On Wed, Dec 7, 2022 at 10:26 PM Aaron Meurer <asmeurer@gmail.com> wrote:
Hi all.
As discussed in today's community meeting, I plan to start working on adding some useful functions to NumPy which are part of the array API standard https://data-apis.org/array-api/latest/index.html.
Although these are all things that will be needed for NumPy to be standard compliant, my focus for now at least is going to be on new functionality that is useful for NumPy independent of the standard. The things that I (and possibly others) plan on working on are:
- A new function matrix_transpose() and corresponding ndarray attribute x.mT. Unlike transpose(), matrix_transpose() will require at least 2 dimensions and only operate on the last two dimensions (it's effectively an alias for swapaxes(x, -1, -2)). This was discussed in the past at https://github.com/numpy/numpy/issues/9530 and https://github.com/numpy/numpy/issues/13797. See
https://data-apis.org/array-api/latest/API_specification/generated/signature...
- namedtuple outputs for eigh, qr, slogdet and svd. This would only apply to the instances where they currently return a tuple (e.g., svd(compute_uv=False) would still just return an array). See the corresponding pages at https://data-apis.org/array-api/latest/extensions/index.html for the namedtuple names. These four functions are the ones that are part of the array API spec, but if there are other functions that aren't part of the spec which we'd like to update to namedtuples as well for consistency, I can look into that.
- New functions matrix_norm() and vector_norm(), which split off the behavior of norm() between vector and matrix specific functionalities. This is a cleaner API and would allow these functions to be proper gufuncs. See
https://data-apis.org/array-api/latest/extensions/generated/signatures.linal...
and https://data-apis.org/array-api/latest/extensions/generated/signatures.linal... .
- New function vecdot() which does a broadcasted 1-D dot product along a specified axis
https://data-apis.org/array-api/latest/API_specification/generated/signature...
- New function svdvals(), which is equivalent to svd(compute_uv=False). The idea here is that functions that have different return types depending on keyword arguments are problematic for various reasons (e.g., they are hard to type annotate), so it's cleaner to split these APIs. Functionality-wise there's not much new here, so this is lower priority than the rest.
- New function permute_dims(), which works just like transpose() but it has a required axis argument. This is more explicit and can't be confused with doing a matrix transpose, which transpose() does not do for stacked matrices by default.
- Adding a copy argument to reshape(). This has already been discussed at https://github.com/numpy/numpy/issues/9818. The main motivation is to replace the current usage of modifying array.shape inplace. (side note: this also still needs to be added to numpy.array_api)
You can see the source code of numpy.array_api for an idea of what pure Python implementations of these changes look like, but to be clear, the proposal here is to add these to the main NumPy namespace, not to numpy.array_api.
One question I have is which of the new functions proposed should be implemented as pure Python wrappers and which should be implemented in C as ufuncs/gufuncs?
Unless there are any objections, I plan to start working on implementing these right away.
Aaron Meurer _______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: ilhanpolat@gmail.com
_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: asmeurer@gmail.com
NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: ilhanpolat@gmail.com