Hi, there is also a stage in the mert training (the actual mert optimization, not running the decoder) that is currently not parallelized, so you cannot expect any gains on that. It is my understanding that we will have a better implementation of mert that also parallelizes the optimization loop, but Nicola could correct me on that.
-phi On Sun, Jun 8, 2008 at 7:52 AM, Yee Seng Chan <[EMAIL PROTECTED]> wrote: > Thanks, increasing to 1G seem to "unstuck" the processes. > My training pairs is only about 250K of sentences, so I had thought 0.5G > should > be sufficient; perhaps the issue is with the size of the LM. > > Nevertheless, I'm doing mert on a dev-set which has only about 1000 source > sentences. For this setup, I found that parallel mert only marginally reduces > the time for each mert iteration. Non-parallel mert needs 60 minutes per > iteration; while parallel-mert on 5 nodes needs 40-50 minutes per iteration. > > I know that the speedup from parallel mert is non-linear in the no. of nodes, > due to the time needed to coordinate among the sub-processes. > > Best regards, > Yee Seng. > > > > > > > _______________________________________________ > Moses-support mailing list > [email protected] > http://mailman.mit.edu/mailman/listinfo/moses-support > > _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
