Dealing with the mode argument in qr.
Hi All,
This post is to bring the discussion of PR #2965https://github.com/numpy/numpy/pull/2965to the attention of the list. There are at least three issues in play here.
1) The PR adds modes 'big' and 'thin' to the current modes 'full', 'r', 'economic' for qr factorization. The problem is that the current 'full' is actually 'thin' and 'big' should be 'full'. The solution here was to raise a FutureWarning on use of 'full', alias it to 'thin' for the time being, and at some distant time change 'full' to alias 'big'.
2) The 'economic' mode serves little purpose. I propose to deprecate it and add a 'qrf' mode instead, corresponding to scipy's 'raw' mode. We can't use 'raw' itself as traditionally the mode may be specified using the first letter only and that leads to a conflict with 'r'.
3) As suggested in 2, the use of single letter abbreviations can constrain the options in choosing mode names and they are not as informative as the full name. A possibility here is to deprecate the use of the abbreviations in favor of the full names.
A longer term problem is the divergence between the numpy and scipy versions of qr. The divergence is enough that I don't see any easy way to come to a common interface, but that is something that would be desirable if possible.
Thoughts?
Chuck
On Tue, Feb 5, 2013 at 2:23 PM, Charles R Harris charlesr.harris@gmail.comwrote:
Hi All,
This post is to bring the discussion of PR #2965https://github.com/numpy/numpy/pull/2965to the attention of the list. There are at least three issues in play here.
- The PR adds modes 'big' and 'thin' to the current modes 'full', 'r',
'economic' for qr factorization. The problem is that the current 'full' is actually 'thin' and 'big' should be 'full'. The solution here was to raise a FutureWarning on use of 'full', alias it to 'thin' for the time being, and at some distant time change 'full' to alias 'big'.
- The 'economic' mode serves little purpose. I propose to deprecate it
and add a 'qrf' mode instead, corresponding to scipy's 'raw' mode. We can't use 'raw' itself as traditionally the mode may be specified using the first letter only and that leads to a conflict with 'r'.
- As suggested in 2, the use of single letter abbreviations can constrain
the options in choosing mode names and they are not as informative as the full name. A possibility here is to deprecate the use of the abbreviations in favor of the full names.
A longer term problem is the divergence between the numpy and scipy versions of qr. The divergence is enough that I don't see any easy way to come to a common interface, but that is something that would be desirable if possible.
Thoughts?
bfroehle has suggested the names
1. complete: Q is a M-by-M matrix, i.e. a complete orthogonal basis. 2. reduced: Q is a M-by-K matrix. 3. r: Only return R 4. raw: Return Householder reflectors Q and TAU
Chuck
On Tue, Feb 5, 2013 at 4:23 PM, Charles R Harris charlesr.harris@gmail.comwrote:
Hi All,
This post is to bring the discussion of PR #2965https://github.com/numpy/numpy/pull/2965to the attention of the list. There are at least three issues in play here.
- The PR adds modes 'big' and 'thin' to the current modes 'full', 'r',
'economic' for qr factorization. The problem is that the current 'full' is actually 'thin' and 'big' should be 'full'. The solution here was to raise a FutureWarning on use of 'full', alias it to 'thin' for the time being, and at some distant time change 'full' to alias 'big'.
- The 'economic' mode serves little purpose. I propose to deprecate it
and add a 'qrf' mode instead, corresponding to scipy's 'raw' mode. We can't use 'raw' itself as traditionally the mode may be specified using the first letter only and that leads to a conflict with 'r'.
- As suggested in 2, the use of single letter abbreviations can constrain
the options in choosing mode names and they are not as informative as the full name. A possibility here is to deprecate the use of the abbreviations in favor of the full names.
A longer term problem is the divergence between the numpy and scipy versions of qr. The divergence is enough that I don't see any easy way to come to a common interface, but that is something that would be desirable if possible.
Thoughts?
Chuck
I would definitely be in favor of deprecating abbreviations.
And while we are on the topic of mode names, scipy.ndimage.filters.percentile_filter() has modes of 'mirror' and 'reflect', and I don't see any documentation stating if they are the same, or what are different about them. I just came across this yesterday.
Cheers! Ben Root
On Wed, Feb 6, 2013 at 3:44 PM, Benjamin Root ben.root@ou.edu wrote:
On Tue, Feb 5, 2013 at 4:23 PM, Charles R Harris < charlesr.harris@gmail.com> wrote:
Hi All,
This post is to bring the discussion of PR #2965https://github.com/numpy/numpy/pull/2965to the attention of the list. There are at least three issues in play here.
- The PR adds modes 'big' and 'thin' to the current modes 'full', 'r',
'economic' for qr factorization. The problem is that the current 'full' is actually 'thin' and 'big' should be 'full'. The solution here was to raise a FutureWarning on use of 'full', alias it to 'thin' for the time being, and at some distant time change 'full' to alias 'big'.
This is asking for problems, to gain some naming consistency. I can't tell how confusing 'full' is now, but deprecating and removing would be better than changing what it returns.
- The 'economic' mode serves little purpose. I propose to deprecate it
and add a 'qrf' mode instead, corresponding to scipy's 'raw' mode. We can't use 'raw' itself as traditionally the mode may be specified using the first letter only and that leads to a conflict with 'r'.
That's not a very good reason to not use "raw", since "raw" is a new option and you therefore don't have to apply the rule that you can give only the first letter to it.
- As suggested in 2, the use of single letter abbreviations can
constrain the options in choosing mode names and they are not as informative as the full name. A possibility here is to deprecate the use of the abbreviations in favor of the full names.
I'm not feeling very strongly about this, but we have to be careful about deprecations. Possible future naming constraints on new modes is not a good reason to deprecate. This one-letter option isn't even mentioned in the docs it looks like. So why not leave that as is and ensure it keeps working (add a unit test if needed)?
A longer term problem is the divergence between the numpy and scipy versions of qr. The divergence is enough that I don't see any easy way to come to a common interface, but that is something that would be desirable if possible.
This would be a problem imho. But I don't see why you can't add "raw" to numpy's qr. And if you add "big" and "thin" in numpy, you can add those modes in scipy too.
Ralf
Thoughts?
Chuck
I would definitely be in favor of deprecating abbreviations.
And while we are on the topic of mode names, scipy.ndimage.filters.percentile_filter() has modes of 'mirror' and 'reflect', and I don't see any documentation stating if they are the same, or what are different about them. I just came across this yesterday.
Cheers! Ben Root
NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
On Wed, Feb 6, 2013 at 12:17 PM, Ralf Gommers ralf.gommers@gmail.comwrote:
On Wed, Feb 6, 2013 at 3:44 PM, Benjamin Root ben.root@ou.edu wrote:
On Tue, Feb 5, 2013 at 4:23 PM, Charles R Harris < charlesr.harris@gmail.com> wrote:
Hi All,
This post is to bring the discussion of PR #2965https://github.com/numpy/numpy/pull/2965to the attention of the list. There are at least three issues in play here.
- The PR adds modes 'big' and 'thin' to the current modes 'full', 'r',
'economic' for qr factorization. The problem is that the current 'full' is actually 'thin' and 'big' should be 'full'. The solution here was to raise a FutureWarning on use of 'full', alias it to 'thin' for the time being, and at some distant time change 'full' to alias 'big'.
This is asking for problems, to gain some naming consistency. I can't tell how confusing 'full' is now, but deprecating and removing would be better than changing what it returns.
That's what the current state of the PR is, both 'full' and 'economic' are deprecated.
- The 'economic' mode serves little purpose. I propose to deprecate it
and add a 'qrf' mode instead, corresponding to scipy's 'raw' mode. We can't use 'raw' itself as traditionally the mode may be specified using the first letter only and that leads to a conflict with 'r'.
That's not a very good reason to not use "raw", since "raw" is a new option and you therefore don't have to apply the rule that you can give only the first letter to it.
Also the current state.
- As suggested in 2, the use of single letter abbreviations can
constrain the options in choosing mode names and they are not as informative as the full name. A possibility here is to deprecate the use of the abbreviations in favor of the full names.
I'm not feeling very strongly about this, but we have to be careful about deprecations. Possible future naming constraints on new modes is not a good reason to deprecate. This one-letter option isn't even mentioned in the docs it looks like. So why not leave that as is and ensure it keeps working (add a unit test if needed)?
Currently qr requires full names for the new modes but not for the deprecated 'full' and 'economic'. That can be changed if we use 'thin' instead of 'reduced'.
A longer term problem is the divergence between the numpy and scipy versions of qr. The divergence is enough that I don't see any easy way to come to a common interface, but that is something that would be desirable if possible.
This would be a problem imho. But I don't see why you can't add "raw" to numpy's qr. And if you add "big" and "thin" in numpy, you can add those modes in scipy too.
Currently I've used bfroehle's suggestions, although I'm tempted by 'thin' instead of 'reduced'
<snip>
Chuck
This would be a problem imho. But I don't see why you can't add "raw" to
numpy's qr. And if you add "big" and "thin" in numpy, you can add those modes in scipy too.
Currently I've used bfroehle's suggestions, although I'm tempted by 'thin' instead of 'reduced'
Thin sounds fine to me. Either way I think we need to clean up the docstring to make the different calling styles more obvious. Perhaps we can just add a quick list of variants: q, r = qr(a) # q is m-by-k, r is k-by-n q, r = qr(a, 'thin') # same as qr(a) q, r = qr(a, 'complete') # q is m-by-n, r is n-by-n r = qr(a, 'r') # r is .... a2 = qr(a, 'economic') # r is contained in the upper triangular part of a2 a2, tau = qr(a, 'raw') # ...
Regards, Brad
On Wed, Feb 6, 2013 at 1:38 PM, Bradley M. Froehle brad.froehle@gmail.comwrote:
This would be a problem imho. But I don't see why you can't add "raw" to
numpy's qr. And if you add "big" and "thin" in numpy, you can add those modes in scipy too.
Currently I've used bfroehle's suggestions, although I'm tempted by 'thin' instead of 'reduced'
Thin sounds fine to me. Either way I think we need to clean up the docstring to make the different calling styles more obvious. Perhaps we can just add a quick list of variants: q, r = qr(a) # q is m-by-k, r is k-by-n q, r = qr(a, 'thin') # same as qr(a) q, r = qr(a, 'complete') # q is m-by-n, r is n-by-n r = qr(a, 'r') # r is .... a2 = qr(a, 'economic') # r is contained in the upper triangular part of a2 a2, tau = qr(a, 'raw') # ...
There is a list similar to that already there. Take a look and comment.
Chuck
participants (4)
-
Benjamin Root -
Bradley M. Froehle -
Charles R Harris -
Ralf Gommers