Kutulutsidwa kwa makina omasulira makina a OpenNMT 0.28.0 (Open Neural Machine Translation), omwe amagwiritsa ntchito njira zophunzirira zamakina, kwasindikizidwa. Kuti apange neural network, polojekitiyi imagwiritsa ntchito luso la laibulale yophunzirira makina yakuya ya TensorFlow. Khodi ya ma module opangidwa ndi pulojekiti ya OpenNMT imalembedwa mu Python ndikugawidwa pansi pa layisensi ya MIT. Zilankhulo zokonzeka zimakonzedwa m'zilankhulo za Chingerezi, Chijeremani ndi Chikatalani; pazilankhulo zina, mutha kupanga paokha chitsanzo kutengera deta kuchokera ku polojekiti ya OPUS (pakuphunzitsidwa, mafayilo awiri amasamutsidwa kudongosolo - imodzi yokhala ndi ziganizo mu chinenero choyambirira, ndipo chachiwiri ndi kumasulira kwapamwamba kwambiri kwa ziganizozi m'chinenero chimene akumasulira ).
Ntchitoyi ikupangidwa mothandizidwa ndi SYSTRAN, kampani yodziwika bwino yopanga zida zomasulira zamakina, ndi gulu la akatswiri ofufuza a Harvard omwe akupanga zitsanzo za zilankhulo za anthu zamakina ophunzirira makina. Mawonekedwe a ogwiritsa ntchito ndi osavuta momwe angathere ndipo amangofunika kufotokozera fayilo yokhala ndi mawu ndi fayilo kuti musunge zotsatira zomasulira. Dongosolo lokulitsa limapangitsa kuti zitheke kugwiritsa ntchito zina zozikidwa pa OpenNMT, mwachitsanzo, kungofotokozera mwachidule, kugawa zolemba komanso kupanga ma subtitle.
Kugwiritsa ntchito TensorFlow kumakupatsani mwayi wogwiritsa ntchito mphamvu za GPU (kufulumizitsa njira yophunzitsira neural network. Kuti muchepetse kugawa kwazinthuzo, polojekitiyi ikupanganso womasulira wodziyimira pawokha mu C ++ - CTranslate2 , yomwe imagwiritsa ntchito zitsanzo zophunzitsidwa kale popanda kutchula zina zowonjezera.
Mtundu watsopanowo umawonjezera parameter yoyambira_learning_rate ndikukhazikitsa mikangano ingapo (mha_bias ndi output_layer_bias) kuti mukonzere jenereta ya Transformer. Zina zonse zimadziwika ndi kukonza zolakwika.
Source: opennet.ru