Hi Jakob There has been a lot of further work on the mira branch, but the changes have not been checked in to the public repository yet. Hopefully this will happen very soon. The latest version produces better results,
Cheers - Barry Sent from my ZX81 ----- Reply message ----- From: "Jakob Elming" <[email protected]> Date: Mon, Aug 8, 2011 12:40 Subject: [Moses-support] Mira optimization branch To: "[email protected]" <[email protected]> Hi, Is anyone working on the mira optimization branch? I have run quite a few tests, and the programs seem to be running ok. The problem is that optimization always leads to a worse bleu score on the tune set that I am optimizing on, compared to the start (unoptimized) parameters. An example: Unoptimized params: BLEU = 12.37, 46.7/16.8/7.7/3.9 (BP=0.998, ratio=0.998, hyp_len=47416, ref_len=47498) First epoch params: BLEU = 11.37, 46.6/16.0/7.1/3.4 (BP=0.982, ratio=0.982, hyp_len=46643, ref_len=47498) Second epoch params: BLEU = 11.35, 46.7/16.0/7.1/3.4 (BP=0.981, ratio=0.981, hyp_len=46605, ref_len=47498) I am optimizing on a vanilla moses system with 14 params being optimized. I have tried different start params with same result. Jakob
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
