Kyauta Mai Ƙara Girman Hoto

Yadda Yake Aiki

Mai ƙara-ƙuduri yana gudanar da Real-ESRGAN general-x4v3 — wani cibiyar sadarwa na convolutional super-resolution na 4.6 MB — ta hanyar onnxruntime-web a cikin wani Web Worker a kan na'urarka. Ana samun model din sau daya daga uwar garkenmu, an ajiye shi, sannan an yi amfani da shi babu haɗi; ana ɓoye hotonka zuwa wani canvas a wurin nan kuma ba a taɓa loda shi ba. Adadin girmatawa tsayayye ne a 4×, saboda an gasa shi kai tsaye a cikin graph a wurare biyu.

1. 1
   Ana yanke hoton zuwa tantani masu haɗuwa

   Sakamakon 4× shine sha shida na pixels, don haka wani hoto na 4000×3000 zai zama megapixels 192 — isasshe don kashe tab ɗin burauza gaba ɗaya maimakon rage shi kawai. Maimakon haka ana raba hoton zuwa cores masu daidaita shi daidai, kuma ana karanta kowace core tare da ƙarin mahallin a kowane gefe don tarin convolution (wanda receptive field ɗinsa yake pixels 34) ya sami ainihin pixels na maƙwabta don yin aiki daga gareshi. Ana yin edge-replicate iyakoki, ba a taɓa yin zero-padding ba, wanda in ba haka ba zai sa model ɗin ya kaifafa wata baki mai baki kewaye da firam.
2. 2
   Ana lulluɓe kowace tantani a matsayin tensor na RGB na planar

   Wata tantani ta zama tensor na float32 na siffar [1, 3, H, W] tare da tashoshin ja, kore da shuɗi an ajiye su a matsayin planes daban kuma ƙimomi kawai an raba su da 255 — babu daidaitawar mean/standard-deviation na kowane iri. Ana rufe kowace tantani zuwa ainihin iri ɗayan girman murabba'i, saboda backend na WebGPU yana sake tara shaders ɗinsa don kowace sabuwar siffar shigarwa kuma wani shiri da manyan tantani uku daban zai biya wannan farashi sau uku.
3. 3
   Inference yana gudana a WebGPU inda take, in ba haka ba a CPU

   Ana ƙirƙirar zama da WebGPU tukuna kuma yana komawa zuwa WebAssembly na zare-guda a duk kasawa — drivers na graphics sau da yawa suna ba da shawarar goyon baya sannan su kasa a tsakiyar ginin graph, kuma hankaka ya fi lalacewa. Saboda madadin yana canja iyakar girma mai gaskiya da lokacin gudu da wani adadi mai girma, kayan aikin yana sake dubawa aikin a kan wanne mai bayarwa ya ainihi samu kafin tantani na farko yayi gudu, kuma yana gaya maka wanne ne ya yi amfani da shi bayan haka.
4. 4
   Ana sake haɗe tantani da cross-fade kuma an daure su

   Haɗe tantani a gefe-da-gefe yana barin wani grid mai ganuwa, don haka ana ba wa gudummawar kowace tantani nauyi ta wani ramp na layi a kan tsagi da suke haɗe. Ramp biyu su ne cikakkun akasi, don haka nauyin suna tarawa daidai zuwa 1 a kowace pixel na fitarwa kuma babu wani layin haske ko duhu da yake bayyana a kan wata seam. Fitarwar danye na model ba a yanke ta ba a cikin graph — ƙimomin da aka auna suna daga kimanin −0.18 zuwa +1.19 — don haka ana daure kowace samfur zuwa zangon 0–255 kafin cast na 8-bit, wanda shine yake hana speckle a kan gefuna masu babban bambanci. Ana sake-samfur alpha daban kuma bilinearly, saboda cibiyar sadarwar da kanta ba ta da hanyar alpha.
5. 5
   An ɓoye sakamakon a cikin worker

   Ana zana pixels da aka haɗe zuwa wani OffscreenCanvas kuma an ɓoye su zuwa PNG nan take, don haka danna Sauke babu gibin async a gabansa. Sauya zuwa JPG yana sake-ɓoyewa daga canvas din da worker din har yanzu yake riƙe da shi maimakon sake gudanar da model din. Wani karamin yanke tsakiya ana kuma zana shi a 100% kusa da iri ɗayan yanki da aka sake-samfur da burauza, wanda shine kaɗai kwatancin da bambancin 4× yake gani da gaske.

Kayan Aiki Masu Alaƙa