Ukusetshenziswa kwesistimu yokufunda yomshini yokuhlanganisa izithombe ngokusekelwe encazelweni yombhalo

Ukuqaliswa okuvulekile kwesistimu yokufunda yomshini i-DALL-E 2, ehlongozwe i-OpenAI, kushicilelwe futhi ikuvumela ukuthi uhlanganise izithombe ezingokoqobo nemidwebo ngokusekelwe encazelweni yombhalo ngolimi lwemvelo, kanye nokusebenzisa imiyalo ngolimi lwemvelo ukuze uhlele izithombe ( isibonelo, engeza, susa noma hambisa izinto ezisesithombeni ). Amamodeli wangempela we-OpenAI we-DALL-E 2 awashicilelwa, kodwa iphepha elinemininingwane yendlela liyatholakala. Ngokusekelwe encazelweni ekhona, abacwaningi abazimele balungiselele okunye ukuqaliswa okubhalwe nge-Python, besebenzisa uhlaka lwe-Pytorch futhi lusatshalaliswa ngaphansi kwelayisensi ye-MIT.

Ukusetshenziswa kwesistimu yokufunda yomshini yokuhlanganisa izithombe ngokusekelwe encazelweni yombhaloUkusetshenziswa kwesistimu yokufunda yomshini yokuhlanganisa izithombe ngokusekelwe encazelweni yombhalo

Uma kuqhathaniswa nokuqaliswa okushicilelwe ngaphambilini kwesizukulwane sokuqala se-DALL-E, inguqulo entsha inikeza ukufana okunembe kakhudlwana kwesithombe encazelweni, ivumela i-photorealism enkulu futhi yenza kube nokwenzeka ukukhiqiza izithombe ngezinqumo eziphakeme. Uhlelo ludinga izinsiza ezinkulu ukuqeqesha imodeli; isibonelo, ukuqeqesha inguqulo yoqobo ye-DALL-E 2 kudinga amahora ayizinkulungwane eziyi-100-200 wekhompyutha ku-GPU, i.e. cishe amaviki angu-2-4 wokubala nge-256 NVIDIA Tesla V100 GPUs.

Ukusetshenziswa kwesistimu yokufunda yomshini yokuhlanganisa izithombe ngokusekelwe encazelweni yombhalo

Umbhali ofanayo naye waqala ukwenza inguqulo enwetshiwe - Ividiyo ye-DALLE2, ehloselwe ukuhlanganisa ividiyo encazelweni yombhalo. Ngokwehlukana, singaqaphela iphrojekthi ye-ru-dalle eyakhiwe yi-Sberbank, ngokuqalisa okuvulekile kwesizukulwane sokuqala se-DALL-E, elungiselelwe ukuqaphela izincazelo ngesiRashiya.

Source: opennet.ru

Engeza amazwana