We often run multiple trainings on the exact same bitext corpus but pull different random samples for each run. We've observed drastically different BLEU scores between different runs with BLEUs ranging from 30 to 45. This is from exactly the same training data except for the randomly-pulled tuning and evaluation sets. We've assumed this difference is due to both the random differences in the sets, floating point variations between various machines and not using --predictable-seeds.
Tom -----Original Message----- From: Hieu Hoang <[email protected]> Reply-to: [email protected] To: John Burger <[email protected]> Cc: Moses-support <[email protected]> Subject: Re: [Moses-support] Nondeterminism during decoding: same config, different n-best lists Date: Thu, 24 Mar 2011 15:51:48 +0000 there's little differences in floating point between OS and gcc versions. One of the regression test fails because of rounding errors, depending on which machine you run it on. Other than truncating the scores, there's not a lot we can do. The mert perl scripts also dabbles in the scores and that may be another source of divergence On 24 March 2011 15:07, John Burger <[email protected]> wrote: Lane Schwartz wrote: > I've examined the n-best lists, and it seems there are at least a > couple of interesting cases. In the simplest case, several > translations of a given sentence produce the exact same score, and > these tied translations appear in different order during different > runs. This is a bit odd, but [not] terribly worrisome. The stranger > case is when there are two different decoding runs, and for a given > sentence, there are translations that appear only in run A, and > different translations that only appear in run B. Both these cases are relevant to something we've occasionally seen, which is non-determinism during =tuning=. This is not surprising given the above, since tuning of course involves decoding. It's hard to reproduce, but we have sometimes seen very different weights coming out of MERT for the exact same system configurations. The problem here is that even very small differences in tuning can result in substantial differences in test results, because of how twitchy BLEU is. Like many folks, we typically run MERT on a cluster. This brings up another source of non-determinism we've theorized about. Some of our clusters are heterogenous, and we've wondered if there might be minor differences in floating point behavior from machine to machine. The assignment of different chunks of the tuning data to different machines is typically non-deterministic, so this might carry over to the actual weights that come out of MERT. Does anyone know how robust the floating point usage in the decoder is under these circumstances? Thanks. - John Burger MITRE _______________________________________________ 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
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
