An buga tsarin fassarar injin na OpenNMT 0.28.0 (Open Neural Machine Translation), wanda ke amfani da hanyoyin koyon inji. Don gina hanyar sadarwa ta jijiyoyi, aikin yana amfani da damar ɗakin karatu mai zurfi na TensorFlow. An rubuta lambar samfuran samfuran da aikin OpenNMT ya haɓaka a cikin Python kuma an rarraba su ƙarƙashin lasisin MIT. An shirya samfuran shirye-shiryen don Ingilishi, Jamusanci da harsunan Catalan; don sauran yarukan, zaku iya ƙirƙirar samfuri da kansa dangane da saitin bayanai daga aikin OPUS (don horarwa, ana canza fayiloli biyu zuwa tsarin - ɗaya tare da jimloli a cikin Harshen tushe, kuma na biyu tare da ingantaccen fassarar waɗannan jimlolin zuwa harshen manufa).
Ana haɓaka aikin tare da haɗin gwiwar SYSTRAN, kamfani mai ƙwarewa wajen ƙirƙirar kayan aikin fassara na'ura, da ƙungiyar masu binciken Harvard waɗanda ke haɓaka ƙirar harshen ɗan adam don tsarin koyon injin. Mai amfani yana da sauƙin sauƙaƙe kamar yadda zai yiwu kuma yana buƙatar kawai ƙayyade fayil ɗin shigarwa tare da rubutu da fayil don adana sakamakon fassarar. Tsarin tsawaita yana ba da damar aiwatar da ƙarin ayyuka bisa OpenNMT, misali, taƙaitawa ta atomatik, rarrabuwar rubutu da ƙarar rubutu.
Yin amfani da TensorFlow yana ba ku damar yin amfani da damar GPU (don hanzarta aiwatar da horo na cibiyar sadarwa na jijiyoyi. Don sauƙaƙe rarraba samfurin, aikin yana haɓaka nau'in mai fassara mai cin gashin kansa a cikin C ++ - CTranslate2). , wanda ke amfani da samfuran da aka riga aka horar ba tare da la'akari da ƙarin abin dogaro ba.
Sabuwar sigar tana ƙara ma'aunin farko_learning_rate kuma yana aiwatar da sabbin muhawara da yawa (mha_bias da fitarwa_layer_bias) don saita janareta na ƙirar Transformer. Sauran ana yiwa alama ta gyaran kwaro.
source: budenet.ru