Buɗe lambar don haɗar motsin rai ta amfani da cibiyoyin sadarwa na jijiya

Ƙungiyar masu bincike daga Jami'ar Fasaha ta Shanghai aka buga kayan aiki Mai kwaikwayo, wanda ke ba da damar yin amfani da hanyoyin koyo na na'ura don kwaikwayi motsin mutane ta amfani da hotuna masu tsattsauran ra'ayi, da kuma maye gurbin tufafi, canza su zuwa wani yanayi kuma canza kusurwar da wani abu yake gani. An rubuta lambar a Python
ta amfani da tsarin PyTorch. Majalisar kuma tana bukata tocila da CUDA Toolkit.

Kayan aikin kayan aiki yana karɓar hoto mai girma biyu azaman shigarwa kuma yana haɗa sakamako da aka gyara dangane da ƙirar da aka zaɓa. Zaɓuɓɓukan canji guda uku ana goyan bayan:
Ƙirƙirar wani abu mai motsi wanda ke biye da motsi wanda aka horar da samfurin. Canja wurin abubuwan bayyanar daga samfurin zuwa abu (misali, canjin tufafi). Ƙirƙirar sabon kusurwa (misali, haɗin hoton bayanin martaba bisa cikakken hoto). Dukkan hanyoyin guda uku za a iya haɗa su, alal misali, za ku iya samar da bidiyo daga hoto wanda ke kwatanta aikin dabarar acrobatic mai rikitarwa a cikin tufafi daban-daban.

A yayin aikin haɗin gwiwa, ana yin ayyukan zaɓin abu a cikin hoto da ƙirƙirar abubuwan da suka ɓace lokacin motsi lokaci guda. Za a iya horar da ƙirar hanyar sadarwa ta jijiyoyi sau ɗaya kuma a yi amfani da su don sauyi daban-daban. Don lodawa akwai shirye-shiryen da aka ƙera waɗanda ke ba ku damar amfani da kayan aikin nan da nan ba tare da horo na farko ba. Ana buƙatar GPU mai girman ƙwaƙwalwar ajiya aƙalla 8GB don aiki.

Ba kamar hanyoyin sauye-sauyen da suka danganci canji ta hanyar mahimman bayanai da ke kwatanta wurin da jiki yake a sararin samaniya mai girma biyu ba, Impersonator yayi ƙoƙari ya haɗa raga mai girma uku tare da bayanin jiki ta amfani da hanyoyin koyo na inji.
Hanyar da aka tsara ta ba da damar yin amfani da manipulations la'akari da keɓaɓɓen siffar jiki da matsayi na yanzu, yin kwaikwayon motsin dabi'a na gabobin.

Don adana bayanan asali kamar laushi, salo, launuka da sanin fuska yayin aiwatar da canji, generative adversarial jijiya cibiyar sadarwa (Farashin GAN). Ana fitar da bayanai game da tushen abu da sigogi don ainihin ganewarsa ta amfani convolutional jijiya cibiyar sadarwa.

source: budenet.ru