Ukukhishwa kwesistimu yokuhumusha yomshini we-OpenNMT 0.28.0 (Open Neural Machine Translation), esebenzisa izindlela zokufunda ngomshini, kushicilelwe. Ukuze kwakhiwe inethiwekhi ye-neural, iphrojekthi isebenzisa amandla omtapo wolwazi wokufunda womshini ojulile we-TensorFlow. Ikhodi yamamojula athuthukiswe iphrojekthi ye-OpenNMT ibhalwe ngePython futhi isatshalaliswa ngaphansi kwelayisensi ye-MIT. Amamodeli enziwe ngomumo alungiselelwa izilimi zesiNgisi, isiJalimane nesiCatalan; kwezinye izilimi, ungakha imodeli ngokuzimela ngokusekelwe kusethi yedatha evela kuphrojekthi ye-OPUS (ukuqeqeshwa, amafayela amabili adluliselwa ohlelweni - elilodwa elinemisho ulimi lomthombo, kanti owesibili nokuhumusha kwekhwalithi ephezulu kwale misho kuya olimini okuqondiwe kulo ).
Le phrojekthi ithuthukiswa ngokubamba iqhaza kwe-SYSTRAN, inkampani ekhethekile ekwakheni amathuluzi okuhumusha ngomshini, kanye neqembu labacwaningi base-Harvard abakha amamodeli olimi lwabantu kumasistimu okufunda emishini. I-interface yomsebenzisi yenziwe lula ngangokunokwenzeka futhi idinga kuphela ukucacisa ifayela lokufaka elinombhalo nefayela ukuze kulondolozwe umphumela wokuhumusha. Isistimu yokwandisa yenza kube nokwenzeka ukuqalisa ukusebenza okwengeziwe okusekelwe ku-OpenNMT, isibonelo, ukufingqa okuzenzakalelayo, ukuhlukaniswa kombhalo nokukhiqizwa kwemibhalo engezansi.
Ukusetshenziswa kwe-TensorFlow kukuvumela ukuthi usebenzise amakhono e-GPU (ukusheshisa inqubo yokuqeqesha inethiwekhi ye-neural. Ukwenza kube lula ukusatshalaliswa komkhiqizo, iphrojekthi iphinde ithuthukise inguqulo ezimele yomhumushi ku-C++ - CTranslate2 , esebenzisa amamodeli aqeqeshwe kusengaphambili ngaphandle kokubhekisela kokuncikile okwengeziwe.
Inguqulo entsha yengeza ipharamitha yokuqalisa_learning_rate futhi isebenzisa izimpikiswano ezintsha ezimbalwa (i-mha_bias kanye ne-output_layer_bias) ukuze kulungiselelwe imodeli ye-Transformer generator. Okunye kumakwa ngokulungiswa kweziphazamisi.
Source: opennet.ru