Re: [Numpy-discussion] Numpy BoF at SciPy 2014 - quick report
Having just re-read the PEP I'm concerned that this proposal leaves at least one major (?) trap for naive users, namely x = np.array([1, 10]) print X.T@x which will print 101, not [[1, 10], [10, 100]]
Yes, I know why this is happening but it's still a problem -- the user said, "I'm thinking matrices" when they wrote @ but the x.T had done the "wrong" thing before the @ kicked in. And yes, a savvy user would have written x = np.ones([[1, 10]]) (but then np.dot(x, x.T) isn't a scalar).
This is the way things are at present, but with the new @ syntax coming in I think we should consider fixing it.
I can think of three possibilities: 1. Leave this as a trap for the unwary, and a reason for people to stick to np.matrix (np.matrix([1, 10]) behaves "correctly") 2. Make x.T a syntax error for 1-D arrays. It's a no-op and IMHO a trap. 3. Make x.T promote the shape == (2,) array to (1, 2) and return a (2, 1) array. This may be too magic, but it's my preferred solution.
R
Implementation of @ (matrix multiplication)
- will be in 3.5 ~ 18months
- no work started yet -- have to make sure we do it.
- @@ was not added.
- The PEP for numpy is well-defined. Not much thinking to be done. (Good for a sprint)
On Do, 2014-07-17 at 09:48 -0400, Robert Lupton the Good wrote:
Having just re-read the PEP I'm concerned that this proposal leaves at least one major (?) trap for naive users, namely x = np.array([1, 10]) print X.T@x which will print 101, not [[1, 10], [10, 100]]
Yes, I know why this is happening but it's still a problem -- the user said, "I'm thinking matrices" when they wrote @ but the x.T had done the "wrong" thing before the @ kicked in. And yes, a savvy user would have written x = np.ones([[1, 10]]) (but then np.dot(x, x.T) isn't a scalar).
This is the way things are at present, but with the new @ syntax coming in I think we should consider fixing it.
I can think of three possibilities:
- Leave this as a trap for the unwary, and a reason for people to stick to np.matrix (np.matrix([1, 10]) behaves "correctly")
- Make x.T a syntax error for 1-D arrays. It's a no-op and IMHO a trap.
- Make x.T promote the shape == (2,) array to (1, 2) and return a (2, 1) array. This may be too magic, but it's my preferred solution.
Making it a warning may be another option. Changing `.T` to promote to 2-d (also maybe to actually only transpose the last two axes for higher D arrays), could be nice, but getting there might take quite a long FutureWarning or even Error -> new feature cycle...
- Sebastian
RImplementation of @ (matrix multiplication)
- will be in 3.5 ~ 18months
- no work started yet -- have to make sure we do it.
- @@ was not added.
- The PEP for numpy is well-defined. Not much thinking to be done. (Good for a sprint)
NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
On Fri, Jul 18, 2014 at 9:03 AM, Sebastian Berg sebastian@sipsolutions.net wrote:
On Do, 2014-07-17 at 09:48 -0400, Robert Lupton the Good wrote:
Having just re-read the PEP I'm concerned that this proposal leaves at least one major (?) trap for naive users, namely x = np.array([1, 10]) print X.T@x which will print 101, not [[1, 10], [10, 100]]
Yes, I know why this is happening but it's still a problem -- the user said, "I'm thinking matrices" when they wrote @ but the x.T had done the "wrong" thing before the @ kicked in. And yes, a savvy user would have written x = np.ones([[1, 10]]) (but then np.dot(x, x.T) isn't a scalar).
This is the way things are at present, but with the new @ syntax coming in I think we should consider fixing it.
I can think of three possibilities: 1. Leave this as a trap for the unwary, and a reason for people to stick to np.matrix (np.matrix([1, 10]) behaves "correctly") 2. Make x.T a syntax error for 1-D arrays. It's a no-op and IMHO a trap. 3. Make x.T promote the shape == (2,) array to (1, 2) and return a (2, 1) array. This may be too magic, but it's my preferred solution.
Making it a warning may be another option. Changing `.T` to promote to 2-d (also maybe to actually only transpose the last two axes for higher D arrays), could be nice, but getting there might take quite a long FutureWarning or even Error -> new feature cycle...
Hmm, just the other day I wrote some code that relies on the current behavior. I was writing a function that could work both on 3-vectors and arrays of 3-vectors. To unpack the input into the separate components, I did:
x, y, z = vector.T
Which works correctly whether `vector` is shaped (3,) or (N, 3).
On Fri, Jul 18, 2014 at 2:03 AM, Sebastian Berg sebastian@sipsolutions.net wrote:
On Do, 2014-07-17 at 09:48 -0400, Robert Lupton the Good wrote:
Having just re-read the PEP I'm concerned that this proposal leaves at
least one major (?) trap for naive users, namely
x = np.array([1, 10]) print X.T@xwhich will print 101, not [[1, 10], [10, 100]]
Yes, I know why this is happening but it's still a problem -- the user
said, "I'm thinking matrices" when they wrote @ but the x.T had done the "wrong" thing before the @ kicked in. And yes, a savvy user would have written x = np.ones([[1, 10]]) (but then np.dot(x, x.T) isn't a scalar).
This is the way things are at present, but with the new @ syntax coming
in I think we should consider fixing it.
I can think of three possibilities: 1. Leave this as a trap for the unwary, and a reason for people to
stick to np.matrix (np.matrix([1, 10]) behaves "correctly")
2. Make x.T a syntax error for 1-D arrays. It's a no-op and IMHOa trap.
3. Make x.T promote the shape == (2,) array to (1, 2) and return a(2, 1) array. This may be too magic, but it's my preferred solution.
Making it a warning may be another option. Changing `.T` to promote to 2-d (also maybe to actually only transpose the last two axes for higher D arrays), could be nice, but getting there might take quite a long FutureWarning or even Error -> new feature cycle...
I've toyed some with the idea of adding a flag bit for transpose of 1-d arrays. It would flip with every transpose and be ignored for non 1-d arrays. A bit of a hack, but would allow for a column/row vector distinction.
Chuck
On Fri, Jul 18, 2014 at 6:18 AM, Charles R Harris <charlesr.harris@gmail.com
wrote:
I've toyed some with the idea of adding a flag bit for transpose of 1-d arrays. It would flip with every transpose and be ignored for non 1-d arrays. A bit of a hack, but would allow for a column/row vector distinction.
very cool! I've thought for a while that one of the major things lacking from numpy.matrix was row and column vectors. To do linear algebra naturally, you really need those.
This may be a really lightweight way to get that - without the distinction between "arrays" and "matrixes", which I think we're trying to get rid of with the @ operator.
when would this flag be used?
- linear algebra operations (mostly @ -- anything else?)
- broadcasting???
neat idea, anyway.
-CHB
On 7/18/2014 4:03 AM, Sebastian Berg wrote:
Changing `.T` to promote to 2-d (also maybe to actually only transpose the last two axes for higher D arrays), could be nice, but getting there might take quite a long FutureWarning or even Error -> new feature cycle.
Considering the extent of implied breakage, I hope this will not be considered. Also, there are already nice ways to add an axis (even optionally, with `atleast_2d`).
I think having `.T` as a no-op on a 1d array is correct behavior. I would not change it. However I can understand preferring an error. (Mathematica considers it an error.)
Alan Isaac
participants (6)
-
Alan G Isaac -
Charles R Harris -
Chris Barker -
Robert Kern -
Robert Lupton the Good -
Sebastian Berg