I-Facebook ithuthukisa i-TransCoder ukuze ihumushe ikhodi isuka kolunye ulimi iye kolunye

Onjiniyela be-Facebook bashicilele i-transcompiler I-TransCoder, esebenzisa amasu okufunda komshini ukuze aguqule ikhodi yomthombo isuke kolunye ulimi lokuhlela lwezinga eliphezulu iye kolunye. Njengamanje, kuhlinzekwa usekelo lokuhumusha ikhodi phakathi kwe-Java, i-C++ ne-Python. Isibonelo, i-TransCoder ikuvumela ukuthi uguqule ikhodi yomthombo we-Java ibe yikhodi ye-Python, kanye nekhodi ye-Python ibe ikhodi yomthombo we-Java. Intuthuko yephrojekthi iyasetshenziswa ucwaningo lwethiyori ekudaleni inethiwekhi ye-neural ukuze kuhlanganiswe ngokuzenzakalelayo okuzenzakalelayo kwekhodi kanye ukubhebhetheka inikezwe ilayisense ngaphansi kwelayisensi ye-Creative Commons Attribution-NonCommerce 4.0 yokusetshenziswa okungezona ezohwebo kuphela.

Ukuqaliswa kwesistimu yokufunda yomshini kusekelwe ku-Pytorch. Amamodeli amabili enziwe ngomumo anikezwa ukuze alandwe: kuqala ukuhunyushwa kwe-C++ ku-Java, i-Java ku-C++ kanye ne-Java ku-Python, kanye okwesibili ezokusakaza
C++ kuya kuPython, Python kuya ku-C++ kanye nePython kuJava. Ukuze siqeqeshe amamodeli, sisebenzise amakhodi omthombo wamaphrojekthi athunyelwe ku-GitHub. Uma uthanda, amamodeli okuhumusha angadalelwa ezinye izilimi zokuhlela. Ukuhlola ikhwalithi yokusakaza, iqoqo lokuhlolwa kweyunithi selilungisiwe, kanye ne-suite yokuhlola ehlanganisa imisebenzi ehambisanayo engama-852.

Kuthiwa ngokunemba kokuguqulwa, i-TransCoder iphakeme kakhulu kunabahumushi bezentengiselwano abasebenzisa izindlela ezisekelwe emithethweni yokuguqulwa, futhi ohlelweni lokusebenza ikuvumela ukuba wenze ngaphandle kokuhlolwa kochwepheshe kochwepheshe kumthombo nolimi okuqondiswe kulo. Amaphutha amaningi avela ngesikhathi sokusebenza kwemodeli angaqedwa ngokungeza imikhawulo elula kusikhiphi sekhodi ukuze kuqinisekiswe ukuthi imisebenzi ekhiqiziwe ilungile ngokwe-syntactically.

Abacwaningi bahlongoze ukwakhiwa okusha kwenethiwekhi ye-neural "Transformer" yokulandelana kokumodela, lapho ukuphindaphinda kungena esikhundleni ngokuthi "ukunaka"(imodeli ye-seq2seq ngokunaka), okukuvumela ukuthi ukhiphe ukuncika okuthile kugrafu yokubala futhi uhambisane nalokho obekungenakulinganiswa ngaphambilini. Zonke izilimi ezisekelwayo zisebenzisa imodeli eyodwa evamile, eqeqeshwa kusetshenziswa izimiso ezintathu—ukuqalisa, ukwenza imodeli yolimi, nokuhumusha emuva.

Source: opennet.ru