Ukukhutshwa kwenkqubo yokuguqulela koomatshini be-OpenNMT 0.28.0 (Open Neural Machine Translation), esebenzisa iindlela zokufunda ngoomatshini, kupapashiwe. Ukwakha uthungelwano lwe-neural, iprojekthi isebenzisa ubunakho bethala leencwadi lokufunda koomatshini obunzulu beTensorFlow. Ikhowudi yeemodyuli eziphuhliswe yiprojekthi ye-OpenNMT ibhalwe kwiPython kwaye isasazwe phantsi kwelayisenisi ye-MIT. Iimodeli ezilungeleyo zilungiselelwe isiNgesi, isiJamani kunye neelwimi zesiCatalan, unokuzimela ngokuzimeleyo imodeli esekelwe kwiseti yedatha evela kwiprojekthi ye-OPUS (uqeqesho, iifayile ezimbini zidluliselwa kwinkqubo - enye enezivakalisi kwi-OPUS; ulwimi lwentsusa, kwaye eyesibini iguqulelwe ngokukumgangatho ophezulu ezi zivakalisi kulwimi ekujoliswe kulo ).
Le projekthi iphuhliswa ngokuthatha inxaxheba kwe-SYSTRAN, inkampani ekhethekileyo ekudaleni izixhobo zokuguqulela ngoomatshini, kunye neqela labaphandi baseHarvard abaphuhlisa imifuziselo yolwimi lwabantu kwiinkqubo zokufunda koomatshini. Ujongano lomsebenzisi lwenziwe lula kangangoko kwaye lufuna kuphela ukucacisa ifayile yegalelo ngombhalo kunye nefayile yokugcina isiphumo sokuguqulela. Inkqubo yolwandiso yenza kube lula ukuphumeza ukusebenza okongeziweyo okusekwe kwi-OpenNMT, umzekelo, ukushwankathela okuzenzekelayo, ukuhlelwa kombhalo kunye nokuveliswa kwemibhalo engezantsi.
Ukusetyenziswa kwe-TensorFlow kukuvumela ukuba usebenzise amandla e-GPU (ukukhawulezisa inkqubo yokuqeqesha inethiwekhi ye-neural. Ukwenza lula ukuhanjiswa kwemveliso, le projekthi iphinda iphuhlise inguqu yokuzimela yomguquleli kwi-C ++ - CTranslate2 , esebenzisa iimodeli eziqeqeshwe kwangaphambili ngaphandle kokubhekisela kwizinto ezixhomekeke kuzo ezongezelelweyo.
Uguqulelo olutsha longeza iparameter yokuqala_learning_rate kwaye iphumeze iingxoxo ezininzi ezintsha (mha_bias kunye ne-output_layer_bias) ukuqwalasela i-Transformer model generator. Intsalela iphawulwe ngolungiso lwebug.
umthombo: opennet.ru