Siphila ezikhathini ezijabulisayo, ezinezimemezelo ezimayelana nobuchwepheshe obuphambili masonto onke. I-OpenAI isanda kukhipha imodeli esezingeni eliphezulu yombhalo-kuya-isithombe i-DALLE 2.
Abantu abambalwa kuphela abathole ukufinyelela kwangaphambi kwesikhathi kusistimu entsha ye-AI engakhiqiza ihluzo ezingokoqobo ezincazelweni zolimi lwemvelo. Kusavaliwe emphakathini.
Uzinzo AI wabe esekhulula i Ukusabalalisa Okuzinzile imodeli, okuhlukile komthombo ovulekile we-DALLE2. Lokhu kwethulwa kushintshe yonke into. Abantu kuyo yonke i-inthanethi bebeshicilela imiphumela esheshayo futhi bemangazwe ubuciko bangempela.
Iyini i-Stable Diffusion?
Ukusabalalisa Okuzinzile imodeli yokufunda yomshini ekwazi ukudala izithombe ezisuka kumbhalo, ukushintsha izithombe kuye ngombhalo, nokugcwalisa imininingwane ezithombeni ezinokulungiswa okuphansi noma ezinemininingwane ephansi.
Iqeqeshwe ngezigidigidi zezithombe futhi ingaletha imiphumela elingana ne I-DALL-E2 futhi MidJourney. Ukuzinza kwe-AI wayisungula, futhi yenziwa esidlangalaleni ngo-Agasti 22, 2022.
Kodwa ngezinsiza ezilinganiselwe zendawo zokubala, imodeli ye-Stable Diffusion ithatha isikhathi eside ukudala izithombe zekhwalithi ephezulu. Ukusebenzisa imodeli ku-inthanethi kusetshenziswa umhlinzeki wamafu kusinikeza izinsiza zokuhlanganisa cishe ezingapheli futhi kusivumela ukuthi sithole imiphumela emihle kakhulu ngokushesha okukhulu.
Ukusingatha imodeli njengesevisi encane kuphinde kuvumela ezinye izinhlelo zokusebenza zokudala ukuthi zisebenzise kalula amandla emodeli ngaphandle kokubhekana nobunzima bokusebenzisa amamodeli e-ML ku-inthanethi.
Kulokhu okuthunyelwe, sizozama ukukhombisa indlela yokuthuthukisa imodeli yokusabalalisa ezinzile futhi siyithumele ku-AWS.
Yakha futhi Ukhiphe Ukusabalalisa Okuzinzile
I-BentoML kanye ne-Amazon Web Services EC2 izindlela ezimbili zokusingatha imodeli ye-Stable Diffusion online. I-BentoML iwuhlaka lomthombo ovulekile lokukala ukufunda imishini amasevisi. Nge-BentoML, sizokwakha isevisi ethembekile yokusabalalisa futhi siyithumele ku-AWS EC2.
Ukulungiselela imvelo nokulanda imodeli yokusabalalisa ezinzile
Izidingo zokufaka futhi uhlanganise inqolobane.
Ungakhetha futhi ulande imodeli ye-Stable Diffusion. Ukunemba okukodwa kulungele ama-CPU noma ama-GPU anobukhulu kuno-10GB we-VRAM. Ukunemba okuyingxenye kulungele ama-GPU ane-VRAM engaphansi kuka-10GB.
Ukwakhiwa Ukusabalalisa Okuzinzile
Sizokwakha isevisi ye-BentoML ukuze sinikeze imodeli ngemuva a I-API ephindaphindiwe. Isibonelo esilandelayo sisebenzisa imodeli yokunemba eyodwa yokubikezela kanye nemojula ye-service.py ukuxhuma isevisi kumqondo webhizinisi. Singadalula imisebenzi njengama-API ngokumaka ngokuthi @svc.api.
Ngaphezu kwalokho, singachaza izinhlobo zokufaka nokukhiphayo zama-API kumapharamitha. Iphoyinti lokugcina le-txt2img, isibonelo, lithola okokufaka kwe-JSON futhi likhiqize okukhiphayo kwesithombe, kuyilapho iphoyinti lokugcina le-img2img lamukela Isithombe nokokufaka kwe-JSON futhi libuyisela okokukhiphayo Kwesithombe.
I-StableDiffusionRunnable ichaza umqondo obalulekile wokukhomba. I-runnable iphethe ukusebenzisa izindlela zamapayipi e-txt2img yemodeli kanye nokuthumela okokufaka okuhlobene. Ukuze usebenzise i-logic inference yemodeli kuma-API, i-Runner yangokwezifiso yakhiwa kusukela ku-StableDiffusionRunnable.
Bese, sebenzisa umyalo olandelayo ukuze uqale isevisi ye-BentoML ukuze ihlolwe. Egijima endaweni Imodeli ye-Stable Diffusion Ukucabanga ngama-CPU kunobuvila. Isicelo ngasinye sizothatha cishe imizuzu emi-5 ukucutshungulwa.
Umbhalo uye esithombeni
Okuphumayo kombhalo kuya esithombeni
Ifayela le-bentofile.yaml lichaza amafayela adingekayo nezinto ezincikile.
Sebenzisa umyalo ongezansi ukuze wakhe i-bento. I-Bento ifomethi yokusabalalisa yesevisi ye-BentoML. Kuyingobo yomlando eziqukethwe ngokwayo equkethe yonke idatha nezilungiselelo ezidingekayo ukuze kuqalwe isevisi.
I-Stable Diffusion bento isiqediwe. Uma ungakwazanga ukwenza kahle i-bento, ungathuki; ungalanda imodeli eyakhiwe ngaphambilini usebenzisa imiyalo esohlwini lwesigaba esilandelayo.
Amamodeli wokwakha ngaphambilini
Okulandelayo amamodeli okwakhiwa ngaphambilini:
Hambisa imodeli ye-Stable Diffusion ku-EC2
Ukuze sikhiphe i-bento ku-EC2, sizosebenzisa i-bentoctl. I-bentoctl ingakuvumela ukuthi usebenzise i-bentos yakho kunoma iyiphi ipulatifomu lamafu usebenzisa i-Terraform. Ukuze wakhe futhi usebenzise amafayela e-Terraform, faka i-opharetha ye-AWS EC2.
Efayelini lokuphakelwa le-config.yaml, ukusetshenziswa sekuvele kulungisiwe. Sicela uzizwe ukhululekile ukuhlela izidingo zakho. I-Bento isetshenziswa ngokuzenzakalelayo kumsingathi we-g4dn.xlarge nge Ukufunda Okujulile I-AMI GPU PyTorch 1.12.0 (Ubuntu 20.04) AMI esifundeni sase-us-west-1.
Dala amafayela e-Terraform manje. Dala isithombe se-Docker bese usilayisha ku-AWS ECR. Kuye ngomkhawulokudonsa wakho, ukulayisha isithombe kungase kuthathe isikhathi eside. Lapho usebenzisa i-bento ku-AWS EC2, sebenzisa amafayela e-Terraform.
Ukuze ufinyelele i-Swagger UI, xhuma kukhonsoli ye-EC2 bese uvula ikheli le-IP lomphakathi esipheqululini. Okokugcina, uma isevisi ye-Stable Diffusion BentoML ingasadingeki, susa ukusetshenziswa.
Isiphetho
Kufanele ukwazi ukubona ukuthi i-SD iheha futhi inamandla kangakanani namamodeli ahambisana nayo. Isikhathi sizosho ukuthi sizophinde siphindaphinde umqondo noma siqhubekele ezindleleni eziyinkimbinkimbi.
Kodwa-ke, okwamanje kunezinhlelo eziqhubekayo zokuqeqesha amamodeli amakhulu anokulungiswa ukuze abambe kangcono indawo ezungezile kanye nemiyalelo. Sizame ukuthuthukisa isevisi ye-Stable Diffusion sisebenzisa i-BentoML futhi sayithumela ku-AWS EC2.
Sikwazile ukusebenzisa imodeli ye-Stable Diffusion kuzingxenyekazi zekhompuyutha ezinamandla kakhudlwana, sakha izithombe ezine-latency ephansi, futhi sanweba ngalé kwekhompyutha eyodwa ngokuphakela isevisi ku-AWS EC2.
shiya impendulo