Kuburitswa kweOpenNMT 0.28.0 (Open Neural Machine Translation) muchina wekushandura sisitimu, inoshandisa nzira dzekudzidza muchina, kwaburitswa. Kuvaka neural network, chirongwa ichi chinoshandisa kugona kweTensorFlow yakadzika muchina raibhurari yekudzidza. Iyo kodhi yemamodule akagadzirwa neOpenNMT purojekiti yakanyorwa muPython uye yakagoverwa pasi peMIT rezenisi. Yakagadzirirwa-yakagadzirwa modhi inogadzirirwa Chirungu, chiGerman uye mitauro yeCatalan; kune mimwe mitauro, unogona kuzvimiririra kugadzira modhi zvichibva pane data set kubva kuOPUS chirongwa (yekudzidziswa, mafaera maviri anoendeswa kune sisitimu - imwe ine mitsara mune mutauro wemabviro, uye yechipiri ine shanduro yemhando yepamusoro yezvirevo izvi mumutauro waunonongedza ).
Chirongwa ichi chiri kuvandudzwa nekubatana kweSYSTRAN, kambani inyanzvi mukugadzira maturusi eshanduro yemuchina, uye boka revatsvagiri veHarvard vari kugadzira mhando dzemitauro yevanhu kuitira masisitimu ekudzidza muchina. Iyo mushandisi interface yakareruka sezvinobvira uye inongoda kutsanangura faira rekuisa rine mavara uye faira kuchengetedza mhinduro yeshanduro. Iyo yekuwedzera sisitimu inoita kuti zvikwanise kuita yekuwedzera mashandiro anoenderana neOpenNMT, semuenzaniso, otomatiki-pfupiso, kurongedza zvinyorwa uye subtitle chizvarwa.
Kushandiswa kweTensorFlow kunokubvumira kushandisa masimba eGPU (kukurumidzira nzira yekudzidzisa neural network. Kuti kurerutsa kugoverwa kwechigadzirwa, chirongwa ichi chiri kugadzirawo shanduro yekuzvimiririra yemuturikiri muC ++ - CTranslate2 , iyo inoshandisa mienzaniso yakadzidziswa isati yataura kune zvimwe zvinotsamira.
Iyo vhezheni itsva inowedzera yekutanga_learning_rate parameter uye inoshandisa akati wandei nharo itsva (mha_bias uye output_layer_bias) kugadzirisa Transformer modhi jenareta. Zvimwe zvacho zvinoratidzwa nebug fixes.
Source: opennet.ru