Details
-
Type:
Story
-
Status: Done
-
Resolution: Done
-
Fix Version/s: None
-
Component/s: faro
-
Labels:None
-
Story Points:1
-
Epic Link:
-
Team:DM Science
-
Urgent?:No
Description
The results from treecorr depend on the underlying OS. I get different answers when running on linux vs macOS. The following snippet shows the issu:
import numpy as np
|
import treecorr
|
import astropy.units as u
|
|
|
np.random.seed(8675309)
|
ra = (np.random.random(100)*0.001 + 2.)*u.radian # Move off 0/360. discontinuity
|
dec = (np.random.random(100)*0.001 - 0.1)*u.radian # Move off equator
|
e1_res = 2*np.random.random(100) - 1.
|
e2_res = 2*np.random.random(100) - 1.
|
|
|
pos_units = 'radian'
|
nbins = 10
|
min_sep = 0.25
|
max_sep = 5.
|
sep_units = 'arcmin'
|
|
|
catTree = treecorr.Catalog(ra=ra, dec=dec, g1=e1_res, g2=e2_res,
|
dec_units=pos_units, ra_units=pos_units)
|
gg = treecorr.GGCorrelation(nbins=nbins, min_sep=min_sep, max_sep=max_sep,
|
sep_units=sep_units)
|
gg.process(catTree)
|
print(gg.meanlogr)
|
On linux this gives:
[-1.2475868 -0.92888378 -0.62015337 -0.32374653 -0.02487182 0.27114931 0.56765978 0.87045272 1.14202255 1.3647294 ]
On macOS, it gives:
[-1.2475868 -0.92888378 -0.62015337 -0.32359399 -0.02516142 0.27028481 0.56836917 0.87014886 1.14075891 1.3647294 ]
Attachments
Activity
A suggestion was made to use `brute=True` for cross platform tests. This should make everything repeatable to machine precision
See the PR for the fix. I'll kick off a jenkins run.
Feel free to pass it on if you don't have the cycles for it
Nice solution - I like having the `brute` option, but not making it the default. Thanks for sorting this out!
Did you actually need to submit a treecorr ticket, or is the behavior sufficiently well understood for that to be unnecessary?
This work was initially accomplished by https://jira.lsstcorp.org/browse/DM-26988 . Following a refactor, the issue was resolved again in https://jira.lsstcorp.org/browse/DM-31611
Adding bin_slop to the call to GGCorrelation helps matching the bins, but it does not completely solve the architecture dependence as the xip values still differ between architectures. This is being dealt with by having reduced precision for that one test, but should be looked into.