{"id":106018,"date":"2022-12-13T09:36:49","date_gmt":"2022-12-13T07:36:49","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/vypusk-sistemy-mashinnogo-perevoda-opennmt-tf-2-30"},"modified":"2022-12-13T09:36:49","modified_gmt":"2022-12-13T07:36:49","slug":"vypusk-sistemy-mashinnogo-perevoda-opennmt-tf-2-30","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/vypusk-sistemy-mashinnogo-perevoda-opennmt-tf-2-30","title":{"rendered":"\u00c7lirimi i sistemit t\u00eb p\u00ebrkthimit automatik OpenNMT-tf 2.30","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
\u00cbllet nj\u00eb version t\u00eb sistemit t\u00eb p\u00ebrkthimit t\u00eb makin\u00ebs OpenNMT-tf 2.30.0 (P\u00ebrkthimi i Makin\u00ebs s\u00eb Nerv\u00ebs), q\u00eb p\u00ebrdor metoda t\u00eb m\u00ebsimit t\u00eb makin\u00ebs. Kodi i moduleve t\u00eb zhvilluara nga projekti OpenNMT-tf \u00ebsht\u00eb shkruar n\u00eb gjuh\u00ebn Python, p\u00ebrdor bibliotek\u00ebn TensorFlow dhe shp\u00ebrndahen n\u00ebn licenc\u00ebn MIT. <\/p>\n
N\u00eb m\u00ebnyr\u00eb paralele, nj\u00eb variant i OpenNMT po zhvillohet mbi bibliotek\u00ebn PyTorch, i cili dallohet n\u00eb nivelin e mund\u00ebsive t\u00eb mb\u00ebshtetura. P\u00ebr m\u00eb tep\u00ebr, OpenNMT mbi PyTorch prezantohet si m\u00eb i leht\u00eb p\u00ebr t'u p\u00ebrdorur dhe multimodal, nd\u00ebrsa varianti mbi TensorFlow shquhet si modular, stabil dhe i aft\u00eb p\u00ebr t\u00eb shfryt\u00ebzuar mund\u00ebsit\u00eb e GPU-s\u00eb p\u00ebr t\u00eb nxitur procesin e m\u00ebsimit t\u00eb rrjetit nervor. P\u00ebr t\u00eb leht\u00ebsuar shp\u00ebrndarjen e produktit, projekti gjithashtu po zhvillon nj\u00eb variant vet\u00eb-mjaftues t\u00eb p\u00ebrkthyesit n\u00eb gjuh\u00ebn C++ \u2014 CTranslate2, i cili p\u00ebrdor modele t\u00eb trajnuara paraprakisht pa u lidhur me var\u00ebsi t\u00eb tjera. <\/p>\n
Modelet jan\u00eb p\u00ebrgatitur p\u00ebr gjuh\u00ebt angleze, gjermane dhe katalanase, nd\u00ebrsa p\u00ebr gjuh\u00ebt e tjera mund t\u00eb formoni vet\u00eb nj\u00eb model bazuar n\u00eb setin e t\u00eb dh\u00ebnave nga projekti OPUS (p\u00ebr t'i m\u00ebsuar sistemit i d\u00ebrgohet nj\u00eb skedar \u2014 nj\u00eb me fraza n\u00eb gjuh\u00ebn origjinale dhe tjetri me p\u00ebrkthimin cil\u00ebsor t\u00eb k\u00ebtyre frazave n\u00eb gjuh\u00ebn e synuar). <\/p>\n
Projekti zhvillohet me pjes\u00ebmarrjen e kompanis\u00eb SYSTRAN, e cila specializohet n\u00eb krijimin e mjeteve p\u00ebrkthimi t\u00eb makin\u00ebs, dhe grupit t\u00eb hulumtuesve t\u00eb Harvardit, q\u00eb zhvillon modele t\u00eb gjuh\u00ebs njer\u00ebzore p\u00ebr sistemet e m\u00ebsimit t\u00eb makin\u00ebs. Nd\u00ebrfaqja e p\u00ebrdoruesit \u00ebsht\u00eb maksimalisht e thjeshtuar dhe k\u00ebrkon vet\u00ebm nj\u00eb skedar hyr\u00ebs me tekstin dhe nj\u00eb skedar p\u00ebr ruajtjen e rezultatit t\u00eb p\u00ebrkthimit. Sistemi i zgjerimeve ofron mund\u00ebsin\u00eb p\u00ebr t\u00eb realizuar funksionalitete t\u00eb tjera mbi OpenNMT, si p\u00ebrkthimi automatik, klasifikimi i teksteve dhe gjenerimi i subtitrave. <\/p>\n
N\u00eb k\u00ebt\u00eb version t\u00eb ri: <\/p>\n
\n
Shtuar mb\u00ebshtetje p\u00ebr bibliotek\u00ebn TensorFlow 2.11, por optimizuesit e rinj Keras ende nuk mb\u00ebshteten (k\u00ebrkohet p\u00ebrdorimi i modalitetit tf.keras.optimizers.legacy).\n
\u00cbsht\u00eb shtuar mb\u00ebshtetje p\u00ebr deg\u00ebn e re t\u00eb motorit CTranslate2 3.x, e cila \u00ebsht\u00eb e destinuar p\u00ebr ekzekutimin efikas t\u00eb modeleve me arkitektur\u00ebn \u00abtransformer\u00bb.\n
Shtuar parametrin e trajnimit t\u00eb modeleve pad_to_bucket_boundary p\u00ebr t\u00eb p\u00ebrfshir\u00eb mbushje shtes\u00eb, duke rrafshuar madh\u00ebsin\u00eb e bllokut n\u00eb vlera shum\u00ebfishin e length_bucket_width.\n
Integruar mb\u00ebshtetje p\u00ebr metrikat chrf dhe chrf++ nga projekti SacreBLEU, q\u00eb krahasojn\u00eb p\u00ebrkthimin e makin\u00ebs me p\u00ebrkthimin e standardizuar njer\u00ebzor.\n
Hequr atributin e modelit ctranslate2_spec, i cili nuk p\u00ebrdoret m\u00eb n\u00eb CTranslate2. <\/ul>\n