-
Notifications
You must be signed in to change notification settings - Fork 134
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
tent/parallel #109
base: master
Are you sure you want to change the base?
tent/parallel #109
Conversation
nno
commented
Feb 25, 2013
- added a module misc/parallelization as a wrapper around whatever kind of parallelization implementation might be available.
- currently supports pprocess and has a fallback single process implementation that should always work.
- refactored Searchlight and surface-based voxel selection code
- added parallelization support to RepeatedMeasure
…ve error messages
…r there are unique file names in case the hdf5 backend is used
Does it pass the full test suite for you -- not just the subset that travis runs? In the past I had problems with serialization of one or the other type that is generated by the test suite. It failed for me once it had to go into another process. |
On 26 February 2013 12:59, Michael Hanke notifications@github.com wrote:
I didn't realize that travis only tests a subset... and the answer is no. (1) Using pprocess and 'pickle' ('native') to store results quite a few (2) Using pprocess and 'hdf5' : FAILED (SKIP=15, errors=5, failures=3) (3) one error when using a single thread process - and this happens not For reference I'm including the errors for each of these cases (note that For those who want to try out these scenarios, look in a) Parallelizer = parallelization.get_best_parallelizer(nproc=1) This sets the number of processes to use to 1, and thus uses a single b) results_backend = 'native' This sets the backend to pickle rather than c) results_backend = Parallelizer.get_best_results_backend() This sets the backend to hdf5 - only effective if nproc!=1 (None or an int In the past I had problems with serialization of one or the other type that
errors using single thread:FAIL: Test AUC computationTraceback (most recent call last): additional errors caused by pprocess+hdf5ERROR: Basic tests of metaclass for using regressions as classifiersTraceback (most recent call last): ERROR: test_split_classifier (mvpa2.tests.test_clf.ClassifiersTests)Traceback (most recent call last): ERROR: Basic tests for TreeClassifierTraceback (most recent call last): ERROR: test_values (mvpa2.tests.test_clf.ClassifiersTests)Traceback (most recent call last): ERROR: mvpa2.tests.test_multiclf.test_multiclass_classifier_cvTraceback (most recent call last): FAIL: test_james_problem (mvpa2.tests.test_rfe.RFETests)Traceback (most recent call last): FAIL: test_james_problem_multiclass (mvpa2.tests.test_rfe.RFETests)Traceback (most recent call last): additional errors caused by pprocess+pickleFAIL: Basic testing of DistPValueTraceback (most recent call last): FAIL: Some really basic testing for match_distributionTraceback (most recent call last): |
Changes Unknown when pulling 7bf68d6 on nno:_tent/parallel into * on PyMVPA:master*. |
Alright, this is another "zombie conversion" ;-) @nno where are you on this one? Is it working for you now? I am increasingly in need of parallelization for RepeatedMeasure, so this would come very handy. |
@Hanke I just updated this branch with the latest master, and things seem to work. That is, unit tests are not failing. The use of parallelization for RepeatedMeasures is switched off for now, as that one is behaving naughty with conditional attributes IIRC. |
I would prefer to leave this one for the next release since new On Thu, 27 Feb 2014, Nikolaas N. Oosterhof wrote:
Yaroslav O. Halchenko, Ph.D. |
@nno I can confirm that it passes for me too. I guess it is expected that the tests do not benefit from any kind of speed-up due to these changes on a multi-core system, right -- all is refactoring.. Would be good to crack the nut on the lost conditional attributes -- can even recall properly what the issue was... |
Assigned milestone 2.4 |
I would be greatly interested by parallelization of RepeatedMeasure. What is the status of this PR? Thanks |