Issue35904

Created on **2019-02-05 23:59** by **rhettinger**, last changed **2019-04-23 08:35** by **rhettinger**. This issue is now **closed**.

Pull Requests | |||
---|---|---|---|

URL | Status | Linked | Edit |

PR 11892 | merged | rhettinger, 2019-02-16 09:14 | |

PR 12919 | merged | rhettinger, 2019-04-23 08:06 |

Messages (19) | |||
---|---|---|---|

msg334894 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-05 23:59 | |

The current mean() function makes heroic efforts to achieve last bit accuracy and when possible to retain the data type of the input. What is needed is an alternative that has a simpler signature, that is much faster, that is highly accurate without demanding perfection, and that is usually what people expect mean() is going to do, the same as their calculators or numpy.mean(): def fmean(seq: Sequence[float]) -> float: return math.fsum(seq) / len(seq) On my current 3.8 build, this code given an approx 500x speed-up (almost three orders of magnitude). Note that having a fast fmean() function is important in resampling statistics where the mean() is typically called many times: http://statistics.about.com/od/Applications/a/Example-Of-Bootstrapping.htm $ ./python.exe -m timeit -r 11 -s 'from random import random' -s 'from statistics import mean' -s 'seq = [random() for i in range(10_000)]' 'mean(seq)' 50 loops, best of 11: 6.8 msec per loop $ ./python.exe -m timeit -r 11 -s 'from random import random' -s 'from math import fsum' -s 'mean=lambda seq: fsum(seq)/len(seq)' -s 'seq = [random() for i in range(10_000)]' 'mean(seq)' 2000 loops, best of 11: 155 usec per loop |
|||

msg334899 - (view) | Author: Josh Rosenberg (josh.r) * | Date: 2019-02-06 02:06 | |

Correct me if I'm wrong, but at least initially, the first listed goal of statistics (per the PEP) was: "Correctness over speed. It is easier to speed up a correct but slow function than to correct a fast but buggy one." numpy already exists for people who need insane speed for these algorithms and are willing to compromise accuracy; am I wrong in my impression that statistics is more about providing correct batteries included that are fast enough for simple uses, not reimplementing numpy piece by piece for hardcore number crunching? Even if such a function were desirable, I don't like the naming symmetry between fsum and fmean; it's kind of misleading. math.fsum is a slower, but more precise, version of the built-in sum. Having statistics.fmean be a faster, less accurate, version of statistics.mean reverses that relationship between the f-prefixed and non-f-prefixed versions of a function. |
|||

msg334902 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-06 04:07 | |

> Correctness over speed. Are you saying that "fsum(seq) / len(seq)" is incorrect because on some older builds there is a rare possibility of an error of 1 unit in the last place? Just about everyone who uses this module will not care one whit about that. ISTM, a requirement for exactness violates "practicality beats purity". We really don't want to be in the position of advising people to never use this module because of design choices that aren't in line with user needs. > It is easier to speed up a correct but slow function > than to correct a fast but buggy one. Taken to an extreme, this could be used justify a BubbleSort. In the case of statistics.mean(), slowness of three orders of magnitude makes this a nearly catastrophic choice for real world use. > I don't like the naming symmetry between fsum and fmean; Any reasonable name will suffice. I suggested an f prefix that could be taken to mean that it is "fast" or that it returns a "float". It isn't much different than isqrt() for an integer square root. > numpy already exists for people who need insane speed for these algorithms > and are willing to compromise accuracy; Let's see what Steven and Tim think about this. |
|||

msg334980 - (view) | Author: Mark Dickinson (mark.dickinson) * | Date: 2019-02-06 20:25 | |

> Are you saying that "fsum(seq) / len(seq)" is incorrect because on some > older builds there is a rare possibility of an error of 1 unit in the last > place? Just to be clear, it's not that rare a possibility, and it's not restricted to older builds. I think Raymond is referring to the bug where on some machines that are still using the x87 for arithmetic, double rounding can lead to the fsum result being out (by potentially an arbitrary amount, not just 1 ulp). But even without that consideration, we'll still often be out by 1 ulp or more on typical new systems. However, if I got my sums right (and assuming that fsum _is_ correctly rounded, and IEEE 754 arithmetic is in use, and the rounding mode hasn't been changed from its default of round-ties-to-even, and we're excluding corner cases like overflow, underflow, etc., etc.), the result of fsum(seq) / len(seq) can never be out by more than 1.5 ulps. That's better than NumPy can promise (even with its use of pairwise summation in some -- but not all -- cases), and should be good enough for almost any practical purpose. I think I'd rather see the regular statistics.mean sacrifice the good-to-the-last-bit accuracy (and it's really not that much of a sacrifice: accurate to 1.5 ulps is a _very_ respectable compromise) than have a new function. I don't know how feasible that is, though, given all the type-handling in statistics.mean. Bikeshedding time: the proposed fmean name seems a bit unfortunate in that it's reversing the sense of sum and fsum, where fsum is the correctly-rounded, slightly slower variant. My first assumption on seeing the name was that fmean was supposed to be to mean as fsum is to sum. |
|||

msg334984 - (view) | Author: Mark Dickinson (mark.dickinson) * | Date: 2019-02-06 20:58 | |

Double-checking my own assertions: here's an example of a list xs of floats for which `fsum(xs) / len(xs)` is out by more than 1 ulp. (Obtained simply by checking a few lists of random.random() outputs; it's probably possible to construct something more obvious.) Python 3.7.2 (default, Dec 30 2018, 08:55:50) [Clang 10.0.0 (clang-1000.11.45.5)] on darwin Type "help", "copyright", "credits" or "license" for more information. >>> import fractions, math>>> xs = [float.fromhex(h) for h in ['0x1.88104e64ffa5cp-3', '0x1.9b793215ddca8p-3', '0x1.754cbf6b09730p-1', '0x1.e2b4ca1df3680p-2', '0x1.91689b66782e1p-1']] >>> approx_mean = math.fsum(xs) / len(xs) >>> approx_mean # in [0.25, 0.5], so 1 ulp is 2**-54 0.47536945341150305 >>> exact_mean = sum(fractions.Fraction(x) for x in xs) / len(xs) >>> exact_mean Fraction(10704368466236809, 22517998136852480) >>> error_in_ulps = abs(exact_mean - fractions.Fraction(approx_mean)) * 2**54 >>> float(error_in_ulps) 1.2 I ran checks on 1000000 such randomly generated lists, and the error never exceeded 1.5 ulps. Sketch of proof of the 1.5 ulps bound: the fsum result is out by at most 0.5 ulps; the length n of the list is presumably represented exactly (lists with more than 2**53 elements would be infeasible). Division of the fsum result by n keeps the relative error the same, but potentially magnifies the ulps error by two, due to the usual "wobble" between relative error and ulps error, so that gives us up to 1 ulp error. Then the result of the division may need to be rounded again, giving another potential error of up to 0.5 ulps. The bound is strict: we can't actually attain 1.5 ulps, so the result we get can't be more than 1 ulp away from the correctly rounded result. > the proposed fmean name seems a bit unfortunate in that it's reversing the sense of sum and fsum I see Josh already made this observation: apologies for the duplicate bikeshedding. |
|||

msg335014 - (view) | Author: Steven D'Aprano (steven.daprano) * | Date: 2019-02-07 12:14 | |

In the PEP, I did say that I was making no attempt to compete with numpy for speed, and that correctness was more important than speed. That doesn't mean I don't care about speed. Nor do I necessarily care about absolute precision when given nothing but float arguments. Mark suggests that using fsum() will be accurate to within 1.5 ulp which satisfies me for float arguments. I doubt that stdev etc would be able to promise that accuracy, so provided your data is all floats, that seems like a pretty good result for the mean. But I'm not super-keen on having two separate mean() functions if that opens the floodgates to people wanting every statistics function to grow a fast-but-floats-only twin. That would make me sad. But maybe mean() is special enough to justify twinning it. In my ideal world, I'd have a single mean() function that had a fast-path for float data, but would automatically drop down to a slower but more accurate path for other types, out-of-range data, etc. I believe that the the built-in sum() function does something like this. When I say "more accurate", this isn't a complaint about fsum(). It refers to the limitation of floats themselves. Call me Mr Silly if you like, but if I need to take the average of numbers bigger than 2**1074 I would like to be able to, even if it takes a bit longer :-) |
|||

msg335045 - (view) | Author: Steven D'Aprano (steven.daprano) * | Date: 2019-02-07 22:34 | |

> def fmean(seq: Sequence[float]) -> float: > return math.fsum(seq) / len(seq) Is it intentional that this doesn't support iterators? |
|||

msg335048 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-08 00:10 | |

>> def fmean(seq: Sequence[float]) -> float: >> return math.fsum(seq) / len(seq) > > Is it intentional that this doesn't support iterators? Since we need both the sum and the length, this seemed like a good starting point. Also, the existing mean() function already covers the more general cases. I suspect that it is common to keep the data in memory so that more than one descriptive statistic can be generated: data = load_measurements() data.sort() n = len(data) mu = fastmean(data) sigma = stdev(data, xbar=mu) low, q1, q2, q3, high = data[0], data[n//4], data[n//2], data[3*n//4], data[-1] popular = mode(data, first_tie=True) It's possible (though possibly not desirable) to provide an fallback path: def fastmean(data: Iterable) -> float: try: return fsum(data) / len(data) except TypeError: # Slow alternative return float(mean(data)) # Memory intensive alternative data = list(data) return fsum(data) / len(data) # Less accurate alternative total = n = 0 for n, x in enumerate(data, start=1): total += x return total / n |
|||

msg335057 - (view) | Author: Steven D'Aprano (steven.daprano) * | Date: 2019-02-08 05:28 | |

> On my current 3.8 build, this code given an approx 500x speed-up On my system, I only get a 30x speed-up using your timeit code. Using ints instead of random floats, I only get a 9x speed-up. This just goes to show how sensitive these timing results are on platform and hardware. What do you think of this implementation? def floatmean(data:Iterable) -> Float: try: n = len(data) except TypeError: # Handle iterators with no len. n = 0 def count(x): nonlocal n n += 1 return x total = math.fsum(map(count, data)) return total/n else: return math.fsum(data)/n Compared to the "no frills" fsum()/len() version: - I see no visible slowdown on lists of floats; - it handles iterators as well. On my computer, the difference between the sequence path and the iterator path is just a factor of 3.5. How does it compare on other machines? As for the name, I think we have three reasonable candidates: float_mean fast_mean fmean (with or without underscores for the first two). Do people have a preference? |
|||

msg335071 - (view) | Author: Steven D'Aprano (steven.daprano) * | Date: 2019-02-08 10:25 | |

Oh, I seem to have accidentally reverted the change of title. Sorry, that was definitely not intended and I don't know how it happened. But now that it has, I'm not going to change it until we have a decision on a name. |
|||

msg335076 - (view) | Author: Mark Dickinson (mark.dickinson) * | Date: 2019-02-08 13:02 | |

No further bikeshedding on the name from me; +1 on whatever you decide on. :-) |
|||

msg335101 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-08 17:56 | |

+1 from me as well. I like your count() solution because 1) it gives the same answer for both an iterator and for an iterable 2) it preserves the memory friendly characteristics of iterators, 3), it is still reasonably fast, and 4) the function signature is still simple. My top name preference is "fmean" because I'm used to "isqrt" for integers, "fsqrt" for floats, and "cmath" for complex where the first letter means the type. Incidentally, that is why "fsum" is named with an "f". My second choice is "fastmean" without an underscore. To my ear, "float_mean" is confusing, as if "float" were a verb and "mean" were the direct object, the antonym of "sink_mean" ;-) |
|||

msg335113 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-08 22:39 | |

Would you like me to submit a PR with docs and tests? |
|||

msg335129 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-09 08:26 | |

FWIW, fabs() and fmod() are other examples of where "f" indicates a function specialized to convert inputs to floats, perform floating point math, and return a float. |
|||

msg335181 - (view) | Author: Steven D'Aprano (steven.daprano) * | Date: 2019-02-10 23:32 | |

> Would you like me to submit a PR with docs and tests? Yes please! I'm happy with the name fmean. |
|||

msg335796 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-18 03:53 | |

> Yes please! I'm happy with the name fmean Okay, the PR is ready. |
|||

msg335877 - (view) | Author: Steven D'Aprano (steven.daprano) * | Date: 2019-02-19 00:01 | |

PR looks good to me, thanks Raymond. Just at the moment I'm having problems with my internet connection leading to technical difficulties with Github. Hopefully I can resolve this soon. |
|||

msg336269 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-02-21 23:06 | |

New changeset 47d9987247bcc45983a6d51fd1ae46d5d356d0f8 by Raymond Hettinger in branch 'master': bpo-35904: Add statistics.fmean() (GH-11892) https://github.com/python/cpython/commit/47d9987247bcc45983a6d51fd1ae46d5d356d0f8 |
|||

msg340702 - (view) | Author: Raymond Hettinger (rhettinger) * | Date: 2019-04-23 08:35 | |

New changeset 7280048690244e73b13f4f68b96c244bcb5434e8 by Raymond Hettinger in branch 'master': bpo-35904: Add missing fmean() entry to the summary table (GH-12919) https://github.com/python/cpython/commit/7280048690244e73b13f4f68b96c244bcb5434e8 |

History | |||
---|---|---|---|

Date | User | Action | Args |

2019-04-23 08:35:18 | rhettinger | set | messages: + msg340702 |

2019-04-23 08:06:33 | rhettinger | set | pull_requests: + pull_request12845 |

2019-02-21 23:07:02 | rhettinger | set | status: open -> closed resolution: fixed stage: patch review -> resolved |

2019-02-21 23:06:36 | rhettinger | set | messages: + msg336269 |

2019-02-19 00:01:47 | steven.daprano | set | messages: + msg335877 |

2019-02-18 03:53:25 | rhettinger | set | messages: + msg335796 |

2019-02-16 09:14:22 | rhettinger | set | keywords:
+ patch stage: patch review pull_requests: + pull_request11920 |

2019-02-10 23:32:48 | steven.daprano | set | messages: + msg335181 |

2019-02-09 08:26:35 | rhettinger | set | messages: + msg335129 |

2019-02-08 22:39:09 | rhettinger | set | messages: + msg335113 |

2019-02-08 17:56:04 | rhettinger | set | messages: + msg335101 |

2019-02-08 13:02:06 | mark.dickinson | set | messages: + msg335076 |

2019-02-08 10:25:54 | steven.daprano | set | messages: + msg335071 |

2019-02-08 05:28:53 | steven.daprano | set | messages:
+ msg335057 title: Add statistics.fastmean(seq) -> Add statistics.fmean(seq) |

2019-02-08 00:16:15 | rhettinger | set | title: Add statistics.fmean(seq) -> Add statistics.fastmean(seq) |

2019-02-08 00:10:58 | rhettinger | set | messages: + msg335048 |

2019-02-07 22:34:48 | steven.daprano | set | messages: + msg335045 |

2019-02-07 12:14:52 | steven.daprano | set | messages: + msg335014 |

2019-02-06 20:58:33 | mark.dickinson | set | messages: + msg334984 |

2019-02-06 20:25:47 | mark.dickinson | set | messages: + msg334980 |

2019-02-06 19:59:04 | mark.dickinson | set | nosy:
+ mark.dickinson |

2019-02-06 04:07:26 | rhettinger | set | messages: + msg334902 |

2019-02-06 02:06:06 | josh.r | set | nosy:
+ josh.r messages: + msg334899 |

2019-02-05 23:59:21 | rhettinger | create |