+1 Sounds like a good plan. On Jun 1, 2013, at 7:54 AM, "Miller, Timothy" <[email protected]> wrote:
> Maybe a good way to do this is to create a big jira and break it up into > module-level jira's, then assign a person to each, and for each module: > 1) Upgrade pom to latest release for modules being used > 2) Recompile/retrain/retest and check in > 3) Close jira for that module > > Then have one last issue that is to > 1) upgrade project-level pom managed versions to latest release > 2) revert module-level poms to use managed versions > 3) run tests and checkin > > How does that plan sound? > Tim > > On 05/30/2013 11:06 PM, Steven Bethard wrote: >> On May 30, 2013, at 8:38 PM, "Masanz, James J." <[email protected]> >> wrote: >>> As a first pass, can we update the dependency to ClearTK 1.4.0 without >>> retraining the models -- would retraining be necessary even if we didn't >>> switch to LIBLINEAR. >> Retraining might still be necessary for anything that uses CleartkExtractor, >> since generation of out-of-bounds features has improved: >> >> https://groups.google.com/forum/#!topic/cleartk-users/KS2BLQcT-ds >> >> But anything not using CleartkExtractor should be safe. And even the things >> that are using CleartkExtractor may not degrade much if the out-of-bounds >> features aren't very important. But we'd need to test it a bit to see. >> >> Steve >> >> >>> Just wonder if we have an option for staging this, in case we can't tackle >>> it all now. >>> >>> -----Original Message----- >>> From: [email protected] >>> [mailto:[email protected]] On Behalf >>> Of Steven Bethard >>> Sent: Thursday, May 30, 2013 5:45 PM >>> To: [email protected] >>> Subject: upgrade ClearTK dependency to 1.4.0? >>> >>> I just released ClearTK 1.4.0 and there are a couple of reasons we should >>> probably consider updating the cTAKES dependency: >>> >>> (1) ClearTK 1.4.0 can now load trained models from the classpath, so we >>> could get rid of the workaround >>> org.apache.ctakes.relationextractor.ae.RelationExtractorAnnotator.allowClassifierModelOnClasspath. >>> >>> (2) ClearTK 1.4.0 has wrappers for multi-class classification with >>> LIBLINEAR which is orders of magnitude faster than LIBSVM. >>> >>> The main downside is that models will have to be re-trained. (It's not >>> necessarily the case that all models would need to be retrained, depending >>> on exactly which classes they were using, but it's probably safer to do so.) >>> >>> I believe this would mostly affect ctakes-temporal, >>> ctakes-relation-extractor and ctakes-assertion. >>> >>> Thoughts? >>> >>> Steve >>> >>> P.S. I noticed that ctakes-assertion declares a dependency on >>> cleartk-examples. The cleartk-examples module was never intended for >>> release, and has not been released as part of ClearTK 1.4.0. Looking at the >>> code, it seems like the dependency in cleartk-examples is not needed, but >>> perhaps a ctakes-assertion person could weigh in on why this dependency was >>> there? >
