Isu tiri kurarama munguva dzinonakidza, nezviziviso nezve yekucheka-kumucheto tekinoroji svondo rega rega. OpenAI ichangoburitsa iyo yekucheka-kumucheto mameseji-kune-mufananidzo modhi DALLE 2.
Vashoma vanhu vakawana mukana wekukurumidza kune itsva AI system inogona kuburitsa echokwadi magiraidhi kubva kurondedzero yemutauro wechisikigo. Ichakavharwa kuruzhinji.
Kugadzikana AI yakabva yaburitsa iyo Yakagadzikana Diffusion modhi, yakavhurika-sosi musiyano weDALLE2. Kuburitswa uku kwakachinja zvese. Vanhu vese painternet vaiburitsa mibairo nekukurumidza uye nekushamiswa nehunyanzvi hwechokwadi.
Chii chinonzi Stable Diffusion?
Yakagadzikana Diffusion imhando yekudzidza muchina inokwanisa kugadzira mifananidzo kubva muzvinyorwa, kushandura mifananidzo zvichienderana nezvinyorwa, uye kuzadza zvakadzama pamifananidzo yakadzika-resolution kana yakaderera-yakadzika.
Yakadzidziswa pamabhiriyoni emifananidzo uye inogona kuunza mhedzisiro yakaenzana ne DALL-E2 uye MidJourney. Kugadzikana AI akaigadzira, uye yakaburitswa pachena muna Nyamavhuvhu 22, 2022.
Asi neane mashoma emunharaunda computational zviwanikwa, iyo Yakagadzikana Diffusion modhi inotora nguva yakareba kugadzira yemhando yepamusoro mapikicha. Kumhanyisa iyo modhi pamhepo uchishandisa gore rinopa kunotipa rinenge risingaperi computational zviwanikwa uye kunotibvumira kuwana yakanakisa mhedzisiro nekukurumidza.
Kugashira modhi se microservice zvakare kunobvumira mamwe maapplication ekugadzira kuti atore nyore nyore kugona kweiyo modhi pasina kutarisana nekuoma kwekumhanyisa ML modhi pamhepo.
Mune ino positi, isu tichaedza kuratidza maitiro ekugadzira yakagadzika diffusion modhi uye nekuiendesa kuAWS.
Vaka uye Deploy Stable Diffusion
BentoML uye Amazon Web Services EC2 inzira mbiri dzekugamuchira iyo Yakagadzikana Diffusion modhi online. BentoML ndeye yakavhurika-sosi sisitimu yekuyera machine learning masevhisi. NeBentoML, tichavaka sevhisi yakavimbika yekuparadzira uye toiendesa kuAWS EC2.
Kugadzirira nharaunda uye kurodha yakagadzika diffusion modhi
Isa zvinodiwa uye gadzira iyo repository.
Iwe unogona kusarudza uye kurodha iyo Yakagadzika Diffusion modhi. Single precision inokodzera maCPU kana maGPU ane yakakura kupfuura 10GB yeVRAM. Hafu chaiyo yakanakira maGPU ane isingasviki 10GB VRAM.
Kuvaka Yakagadzikana Diffusion
Tichavaka sevhisi yeBentoML yekushandira iyo modhi kuseri kwe RESTful API. Muenzaniso unotevera unoshandisa iyo imwechete chaiyo modhi yekufembera uye iyo service.py module kubatanidza sevhisi kune bhizinesi pfungwa. Tinogona kufumura mabasa semaAPI nekuvamaka ne @svc.api.
Uyezve, isu tinokwanisa kutsanangura iyo APIs 'yekupinza uye inobuda mhando mumaparamita. Iyo txt2img endpoint, semuenzaniso, inogamuchira yeJSON yekupinda uye inoburitsa Mufananidzo wakabuda, nepo img2img endpoint inogamuchira Mufananidzo uye JSON yekupinda uye inodzosa Mufananidzo wakabuda.
A StableDiffusionRunnable inotsanangura yakakosha inference logic. Iyo inomhanya inotungamira yekumhanyisa iyo modhi txt2img pombi nzira uye kutumira mune akakodzera mapimendi. Kumhanyisa iyo inference inference logic mumaAPIs, tsika Runner inovakwa kubva kuStableDiffusionRunnable.
Zvadaro, shandisa murairo unotevera kutanga sevhisi yeBentoML yekuedza. Locally running the Yakagadzikana Diffusion modhi kufungidzira paCPUs pane husimbe. Chikumbiro chega chega chichatora anenge maminetsi mashanu kuti chigadziriswe.
Runyoro kumufananidzo
Runyoro kuenda kumufananidzo
Iyo bentofile.yaml faira inotsanangura mafaera anodiwa uye zvinoenderana.
Shandisa murairo pazasi kuvaka bento. Bento ndiyo nzira yekugovera yebasa reBentoML. Iyo inochengeterwa yega iyo ine data rese uye zvigadziriso zvinodiwa kuti utange sevhisi.
Iyo Stable Diffusion bento yakapedzwa. Kana iwe usina kukwanisa kugadzira zvakanaka bento, usavhunduka; unogona kudhawunirodha pre-yakavakwa modhi uchishandisa mirairo yakanyorwa muchikamu chinotevera.
Pre-kuvaka Models
Zvinotevera ndizvo pre-kuvaka modhi:
Shandisa Yakagadzikana Diffusion modhi kuEC2
Kuendesa bento kuEC2, isu tichashandisa bentoctl. bentoctl inogona kukutendera kuti uendese mabentos ako kune chero gore chikuva kushandisa Terraform. Kuvaka uye kushandisa mafaira eTerraform, isa iyo AWS EC2 opareta.
Mune deployment config.yaml faira, kutumirwa kwakatogadzirwa. Ndapota inzwa wakasununguka kugadzirisa kune zvaunoda. Bento inoiswa nekusarudzika pane g4dn.xlarge host ine Deep Learning AMI GPU PyTorch 1.12.0 (Ubuntu 20.04) AMI pane us-kumadokero-1 dunhu.
Gadzira mafaira eTerraform izvozvi. Gadzira iyo Docker mufananidzo uye uiise kuAWS ECR. Zvichienderana nebandwidth yako, kukwirisa mufananidzo kunogona kutora nguva yakareba. Pakuisa bento kuAWS EC2, shandisa Terraform mafaira.
Kuti uwane iyo Swagger UI, batanidza kune EC2 koni uye vhura iyo yeruzhinji IP kero mubrowser. Pakupedzisira, kana Stable Diffusion BentoML sevhisi isingachadiwi, bvisa kutumirwa.
mhedziso
Iwe unofanirwa kuona kuti inonakidza uye ine simba sei SD uye inofambidzana nayo mhando. Nguva ichatiudza kana isu tichizodzokorora pane iyo pfungwa mberi kana kuenderera kune dzimwe nzira dzakaomarara.
Zvakadaro, pari zvino pane zvirongwa zviri kuitika zvekudzidzisa mamodheru makuru ane zvigadziriso kuti zvinyatsobata zvakatenderedza uye mirairo. Takaedza kugadzira iyo Stable Diffusion sevhisi tichishandisa BentoML uye tikaiendesa kuAWS EC2.
Isu takakwanisa kumhanyisa iyo Yakagadzikana Diffusion modhi pane yakanyanya simba Hardware, kugadzira mifananidzo ine yakaderera latency, uye kuwedzera kupfuura komputa imwe chete nekuisa sevhisi paAWS EC2.
Leave a Reply