Opened 11 years ago

Closed 11 years ago

#12558 closed defect (fixed)

Numerical noise in hmm.pyx

Reported by: Volker Braun Owned by: Minh Van Nguyen
Priority: blocker Milestone: sage-5.0
Component: doctest coverage Keywords: hidden markov model
Cc: mhampton, Jeroen Demeyer Merged in: sage-5.0.beta6
Authors: Volker Braun Reviewers: Jeroen Demeyer
Report Upstream: N/A Work issues:
Branch: Commit:
Dependencies: Stopgaps:

Status badges

Description

I found some more numerical noise in this doctest in the hidden markov model stuff:

File "/home/vbraun/tesla/sage-5.0.beta4/devel/sage/sage/stats/hmm/hmm.pyx", line 281:
    sage: m
Expected:
    Discrete Hidden Markov Model with 2 States and 2 Emissions
    Transition matrix:
    [1.0134345614...e-70               1.0]
    [              1.0 3.99743527136e-19]
    Emission matrix:
    [7.3802215662...e-54               1.0]
    [              1.0  3.9974352626e-19]
    Initial probabilities: [0.0000, 1.0000]
Got:
    Discrete Hidden Markov Model with 2 States and 2 Emissions
    Transition matrix:
    [1.01343456147e-70               1.0]
    [              1.0 3.99743527135e-19]
    Emission matrix:
    [7.38022156625e-54               1.0]
    [              1.0  3.9974352626e-19]
    Initial probabilities: [0.0000, 1.0000]

Attachments (1)

trac_12558_numerical_noise_in_hmm.patch (897 bytes) - added by Volker Braun 11 years ago.
Initial patch

Download all attachments as: .zip

Change History (7)

Changed 11 years ago by Volker Braun

Initial patch

comment:1 Changed 11 years ago by Volker Braun

Authors: Volker Braun
Cc: mhampton Jeroen Demeyer added
Priority: majorblocker
Status: newneeds_review

comment:2 Changed 11 years ago by Jeroen Demeyer

Reviewers: Jeroen Demeyer
Status: needs_reviewpositive_review

Could you state on which system this happens? The patch obviously looks good.

comment:3 Changed 11 years ago by Volker Braun

This is on a i7-920, 12GB ram, Tesla GPGPU card (of course not used in the doctest), running Ubuntu 11.10 i386(!). Taking the less beaten path with a 32-bit OS (don't ask).

comment:4 in reply to:  3 Changed 11 years ago by Jeroen Demeyer

Replying to vbraun:

Tesla GPGPU card (of course not used in the doctest)

would be über-cool though...

running Ubuntu 11.10 i386(!). Taking the less beaten path with a 32-bit OS (don't ask).

Interesting that I haven't seen this failure on Skynet cicero (Fedora 15 i386, Pentium 4).

comment:5 Changed 11 years ago by Volker Braun

Cicero is running Fedora 16, for the record.

The i7 has more SSE ops, if the compiler used those instead of FPU then that would easily explain the slightly different precision. GCC on tesla is 4.6.1.

comment:6 Changed 11 years ago by Jeroen Demeyer

Merged in: sage-5.0.beta6
Resolution: fixed
Status: positive_reviewclosed
Note: See TracTickets for help on using tickets.