State takes 4 + 8 * (kMaxOrder - 1) bytes per hypothesis. It does not impact LM size (except of course you're loading a larger LM), just hypotheses. To be honest, I don't have a good way of telling you how many hypotheses there are; suppose you could increase this and measure the difference. In practice, hypothesis memory consumption is rather small compared to the phrase table and language model. I could work on shimming a runtime state size into the existing memory allocation that Moses does for each hypothesis. However, people typically run with models up to order 6, so this is somewhat lower priority.
Also, I think you'll like the 23% speed improvement for trie that I checked in as revision 3988. Kenneth On 05/24/11 01:07, Tom Hoar wrote: > Thanks. > > If I change the value to something like 12, how does this affect the > overall memory requirements/allocation for the State object? I.e. if I > compile with kMaxOrder = 12, and use Kenlm for a model with order = 6, > is more memory required/allocated and if so, how much? Or, does the > additional allocation only occur when the model has a higher order? > > Tom > > > > On Mon, 23 May 2011 16:33:01 -0400, Kenneth Heafield > <[email protected]> wrote: >> Edit kenlm/lm/max_order.hh and recompile. >> >> The reason is to minimize the size of the State object held by each >> hypothesis while avoiding dynamic memory allocation. >> >> On 05/23/11 15:39, Tom Hoar wrote: >>> I use KenLM's build_binary for language models. There are no problems >>> order values up to 6 grams. It halts 7 grams with a warning to change a >>> value somewhere to support higher values. I don't have a current log >>> output. Can you share what change to make and where? >>> >>> Thanks. >>> Tom >>> >>> >>> >>> _______________________________________________ >>> 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
