Ok, thank you Marcin. On Fri, May 10, 2019 at 11:53 AM Marcin Junczys-Dowmunt <[email protected]> wrote:
> Hi, > > Yes, a smaller phrase table should help. I wrote the table, but that was > in 2012 and I cannot really remember what goes on in there. I think making > sure that you do not have too many target phrases per source phrase should > help. > > > > *From: *He Shiming <[email protected]> > *Sent: *Thursday, May 9, 2019 8:49 PM > *To: *[email protected] > *Subject: *[Moses-support] processPhraseTableMin Cannot encode numbers > largerthan 268435455 > > > > Hi, > > > > I'm training a Chinese-to-English phrase-based model, using 33 million > sentence pairs. My phrase table is 90GB gzipped, and the reordering table > is 27GB gzipped. When running processPhraseTableMin, it dies in step 3 > because of the following error: > > > > Intermezzo: Calculating Huffman code sets > > Creating Huffman codes for 1786817 target phrase symbols > > Creating Huffman codes for 871265 scores > > Creating Huffman codes for 18018117 scores > > Creating Huffman codes for 827039 scores > > Creating Huffman codes for 17861459 scores > > Creating Huffman codes for 50 alignment points > > > > Pass 3/3: Compressing target phrases > > ..................................................[5000000] > > ..................................................[345000000] > > ............................................terminate called after > throwing an instance of 'util::Exception' > > what(): moses/TranslationModel/CompactPT/ListCoders.h:179 in static > void Moses::Simple9::EncodeSymbol(Moses::Simple9::uint&, InIt, InIt) [with > InIt = unsigned int*; Moses::Simple9::uint = unsigned int] threw > util::Exception because `*it > 268435455'. > > You are trying to encode 436766721 with Simple9. Cannot encode numbers > larger than 268435455 (2^28-1) > > Aborted (core dumped) > > > > Is my phrase table too big? Pruning seems to have only removed 0.1% of the > phrases. Is retraining using fewer pairs my only option? > > > > -- > > Best regards, > He Shiming > > > -- Best regards, He Shiming
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
