Hi, I still have problems with getting a factored model on Europarl data to run without memory allocation errors. I tried a simpler setup now with 3 decoding steps, 2 translation steps and one generation step. Translations are done on lemmas and POS labels and the generation step is supposed to create surface forms from target lemmas and POSs. Both phrase tables and the reordering table are binarized and the LM is also in binary format using IRST. Caching is disabled. However, moses still fails with a memory allocation error after some sentences (about 100) using the non-tuned model. Surprisingly this happens with a very short sentence of only 8 words. There's nothing really strange with the sentence either (the only thing is a hyphen in the beginning but removing it doesn't change anything).
Now I'm wondering if I did something wrong or if europarl is just too big for such experiments. What would be a typical factored model that could be run on this amount of data? I'm running on a server with 8GB RAM - so that shouldn't be the big problem. Another question: would it be possible to run a generation step (for example lemma+POS --> word) with a second model instead of doing it in the main decoding step (in the same way as the recaser is done afterwards)? Would that be a solution for my memory problems? If this is possible should I expect much different (worse) results? thanks in advance for any suggestions, Jörg _______________________________________________ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support
