Avoid numerical instability #1652
Conversation
This avoids basically doing 1 - 1, for example: >>> from math import exp >>> margin = -40 >>> 1 - 1 / (1 + exp(margin)) 0.0 >>> exp(margin) / (1 + exp(margin)) 4.248354255291589e-18 >>>
|
Can one of the admins verify this patch? |
|
Y'know, there's a similar issue in
For -40, this gives 0.0, when really it should be about math.exp(-40) = 4.248354255291589e-18, since log(1+x) ~= x for very small x. This one can be fixed up with
I'll have a look for other instances beyond the 4 I see and open a JIRA? I could mention this PR too to bring it under one umbrella. |
|
See also https://issues.apache.org/jira/browse/SPARK-2748 and #1659 . This could be considered part of SPARK-2748. |
|
Jenkins, add to whitelist. |
|
Jenkins, test this please. |
|
QA tests have started for PR 1652. This patch merges cleanly. |
…Math.exp, Math.log In a few places in MLlib, an expression of the form `log(1.0 + p)` is evaluated. When p is so small that `1.0 + p == 1.0`, the result is 0.0. However the correct answer is very near `p`. This is why `Math.log1p` exists. Similarly for one instance of `exp(m) - 1` in GraphX; there's a special `Math.expm1` method. While the errors occur only for very small arguments, given their use in machine learning algorithms, this is entirely possible. Also note the related PR for Python: #1652 Author: Sean Owen <srowen@gmail.com> Closes #1659 from srowen/SPARK-2748 and squashes the following commits: c5926d4 [Sean Owen] Use log1p, expm1 for better precision for tiny arguments
|
QA results for PR 1652: |
|
LGTM. Merged into master. Thanks! |
|
Awesome, thank you! :-) |
…Math.exp, Math.log In a few places in MLlib, an expression of the form `log(1.0 + p)` is evaluated. When p is so small that `1.0 + p == 1.0`, the result is 0.0. However the correct answer is very near `p`. This is why `Math.log1p` exists. Similarly for one instance of `exp(m) - 1` in GraphX; there's a special `Math.expm1` method. While the errors occur only for very small arguments, given their use in machine learning algorithms, this is entirely possible. Also note the related PR for Python: apache#1652 Author: Sean Owen <srowen@gmail.com> Closes apache#1659 from srowen/SPARK-2748 and squashes the following commits: c5926d4 [Sean Owen] Use log1p, expm1 for better precision for tiny arguments
This avoids basically doing 1 - 1, for example: ```python >>> from math import exp >>> margin = -40 >>> 1 - 1 / (1 + exp(margin)) 0.0 >>> exp(margin) / (1 + exp(margin)) 4.248354255291589e-18 >>> ``` Author: Naftali Harris <naftaliharris@gmail.com> Closes apache#1652 from naftaliharris/patch-2 and squashes the following commits: 0d55a9f [Naftali Harris] Avoid numerical instability
5 participants
This avoids basically doing 1 - 1, for example: