Cogo min na pusi ye diffusion sabatilen debug ( 2023 )

AI tariku kɔnɔ, bɔnɛko maana minnu tun ma labɛn fɔlɔ, olu dɔ la kelen kɔnɔ, so kɔnɔ pusi dɔ ye ɛntɛrinɛti injiniyɛriw dɛmɛ k’a sɔrɔ u ma a dɔn, ka yɔrɔ dogolen tiɲɛni jugu dɔ dɔn Stable Diffusion ka ja dilancogo pibiliki kɔnɔ. San 2023 ko in kɛra ko kɛlen ye min kɛra fɛnba ye, min b’a jira cogo min na, diɲɛ kɔnɔfɛnw dontaw minnu tɛ se ka fɔ ka ɲɛ, olu bɛ se ka filiw bɔ kɛnɛ kan, minnu bɛ tɛmɛ pewu lɛrɛ ba caman kɔnɔ, sɛgɛsɛgɛli sigilenw kɔnɔ.

Mun kɛra tiɲɛ na ni pusi ni jɛnsɛnni sabatilen ye ?

san 2023 daminɛ na , masin kalan injiniyɛri dɔ min tun bɛ baara kɛ ka bɔ so , o ye fɛn kɛrɛnkɛrɛnnen dɔ kɔlɔsi . U ka pusi, n’a taamana ka tɛmɛn klaviyeti kan Stable Diffusion degeliw boli senfɛ, o ye mɔgɔ hakilintanw kulu dɔ don teliya la kulu dɔ kɔnɔ. Sani a ka bɔli nɔgɔlenw bɔ walima ka fili dɔ fili, modɛli ye ja dɔw labɛn ni yecogo fɛn ye min bɛ bɛn ɲɔgɔn ma ani min kɛrɛnkɛrɛnnen don kosɛbɛ — tessellation pattern segin-segin min tun man kan ka kɛ ni donna teliyalenw dira.

nin tun tɛ mankan gansan ye . O misali in ye bias (kɔrɔbaliya) dɔ jira min tun ma dɔn fɔlɔ modɛli ka cross-attention layers kɔnɔ, kɛrɛnkɛrɛnnenya la, U-Net architecture (U-Net architecture) ye taamasiyɛn dɔw faralen ɲɔgɔn kan cogo min na, minnu binna kanko dancɛ nɔgɔmanw kɔfɛ. Pusi ka klaviyeti mashing tun ye kɛlɛ-ɲini-minɛn dɔ Dabɔ koɲuman, hadamaden kɔrɔbɔla si tun ma miiri min na k’a lajɛ, o ye fili dɔ jira modeli in ka CLIP sɛbɛnni encodeur integration la min ye nɔ bila yɔrɔko jɛɲɔgɔnyaw jatebɔcogo la denoising kɛcogo la.

Enjiniyɛri jɛkulu ye dɔgɔkun nataw kɛ ka fɛn dilannen in nɔfɛtaama ka segin a ju la : ko dɔ ye min bɛ wuli-yɔrɔ laminikogo la latent diffusion scheduler kɔnɔ min bɛ jira dɔrɔn tokenization edge cases kɛrɛnkɛrɛnnenw jukɔrɔ. O labɛnni ye ja ɲɔgɔndan ɲɛ ka taa a fɛ prompt suguya bɛɛ la ni 3-4% ye, o ye tɔnɔba ye AI baarakɛcogo la min bɛ bɔ mɔgɔw la.

Mun na donnakow minnu tɛ laadalako ye, olu bɛ buguw minɛ minnu bɛ sɔrɔ QA ekipuw fɛ ?

Sɔrɔbɔli sigilen bɛ tugu hadamaden ka hakilinata kɔ. Ɛntɛrinɛti ɲɛmɔgɔw bɛ sɛgɛsɛgɛli kɛcogo sɛbɛn ka da baarakɛlaw ka kɛwale makɔnɔnenw kan, dakun ko minnu bɛ se ka miiri u la, ani dɛsɛ cogoya dɔntaw ka bɔ iterasiyɔn tɛmɛnenw na. Nka porogaramuw — kɛrɛnkɛrɛnnenya la AI sitɛmu minnu bɛ ni paramɛtiri miliyari caman ye — olu kɔnɔ, jamana minnu bɛ se ka kɛ, olu ka bɔnɔgɔla faralen ɲɔgɔn kan, kɔrɔbɔli hukumu si tɛ se ka min datugu pewu.

ye

"Bug minnu farati ka bon kosɛbɛ, olu tɛ minnu dogolen don kode kɔnɔ i ma minnu kɔrɔbɔ. U ye minnu dogolen don kode kɔnɔ i ye minnu kɔrɔbɔ ni hakilina juguw ye." — Nin sariyasen in, min faamuyara kabini tuma jan laadala porozɛw dilanni na, o bɛ kɛ fɛnba ye kosɛbɛ masin kalanni siratigɛ la, donna yɔrɔ min na, dan tɛ yɔrɔ minnu na.

Pusi ko in ye fanga di ɲagami injiniyɛri baarakɛlaw ye min dɔn kabini san caman : donnakow minnu kɛra k’a sɔrɔ u ma kɛ cogo si la, minnu tɛ se ka fɔ ka ɲɛ, olu bɛ sigida barikantanya jira, fɛɛrɛ sɛgɛsɛgɛli tɛ se ka minnu kɛ. O sariyakolo kelen de bɛ fuzz testing kɔfɛ, n’o ye kunnafonidilanw ye minnu ma labɛn ka ɲɛ ni laɲini ye, olu bɛ don sistɛmuw kɔnɔ walasa ka dɛsɛw bɔ kɛnɛ kan. danfara min bɛ yan , o ye ko fusi sen naani ni kùnkolo tun b' a la .

O ye mun jira AI debugging gɛlɛyaw kan ?

AI modeli generatifw debugging ni laadala porogaramuw debugging tɛ kelen ye kosɛbɛ. Ni baarakɛminɛn kɔrɔ dɔ Dɛsɛra, i bɛ fili sɛbɛn Sɔrɔ, stack trace, sira min bɛ Se ka Lasegin. Ni AI modɛli dɔ ye bɔli juguw bɔ cogo nɔgɔman na, o dɛsɛ bɛ se ka kɛ kalo caman kɔnɔ k’a sɔrɔ u ma kɔlɔsi bawo jaabi "tiɲɛnen" kelen tɛ yen min bɛ se ka suma ni o ye.

• Yɔrɔ dogolen opacity : kɔnɔna jiracogo minnu bɛ diffusion models kɔnɔ, olu tɔgɔ bɔra kosɛbɛ ka kɔrɔfɔ ka gɛlɛn, o b’a to a ka gɛlɛn ka bɔli kɛlenw nɔfɛtaama ka segin jatebɔ dɛsɛ kɛrɛnkɛrɛnnenw kan.
• Sɔnni teliya la : Yɛlɛma misɛnnin minnu bɛ sɛbɛnni doncogo la , olu bɛ se ka bɔli suguya wɛrɛw lase kungo kɔnɔ , o kɔrɔ ye ko buguw bɛ se ka bɔ kɛnɛ kan dɔrɔn cogoya fitininw na minnu tɛ se ka fɔ ka ɲɛ .
• Jateminɛ kɛcogo : A tɛ i n’a fɔ danfara baara minnu bɛ kɛ ni tiɲɛni ye min bɛ se ka suman , ja dilancogo ɲuman yɔrɔ dɔ ye hakilijagabɔ ye , o b’a to tiɲɛni misɛnninw bɛ se ka sɛgɛn otomatiki sɛgɛsɛgɛliw fɛ .
• Kaskade dependencies : Nafa kelen min bɛ sɛbɛnni kode la, o bɛ se ka jɛnsɛn cross-attention mechanism fɛ, denoising scheduler fɛ, ani VAE decoder fɛ, o bɛ kɛ sababu ye ka root cause analysis kɛ gɛlɛya ye kosɛbɛ.
• Kalan kunnafoniw sirili : Danfara kɛli buguw cɛ modeli jɔcogo la ani bias minnu bɛ ciyɛn sɔrɔ kalan kunnafoniw na , o bɛ ablation kalanw de wajibiya minnu bɛ waati ta ani minnu bɛ jatebɔ musaka bɔ .

Nin ko in ye nɔ bila AI yiriwali walew la cogo di ?

Cusi debugging maana , hali n' a ye nisɔndiyako ye a ɲɛda la , o ye fɛn caman wuli fɛn caman na AI ekipuw ka jogoɲumanya sabatili siratigɛ la . Kabini o waati, jɛkulu caman y’u ka fuzz testing protocoles (fuzz testing protocols) bonya ka ɲɛsin generative models ma, kɛrɛnkɛrɛnnenya la, u bɛ token sequences randomes ani adversarial token sequences (taamaʃyɛnw) sigi senkan minnu bɛ donnakow ladege minnu tɛ kanko ye. Sisan, ekipu dɔw bɛ "klaviyeti taama" simulatigɛ otomatikiw kɛ u ka jɛ-ka-baara pibilikiw kɔnɔ minnu bɛ taa ɲɛ.

O ko in ye nafa kura sɔrɔ fana faamuyali baarakɛminɛnw na diffusion models kama . Ni o fɛn yelen tun ma ye kosɛbɛ — kulɛri caman cili nɔgɔlen sanni ka kɛ tessellation ja gɛlɛn ye — a tun bɛ se ka kɛ ko a ma kɔlɔsi fo abada. O ye sigidamɔgɔw bila ka taa anomali dɔnni otomatiki ɲuman labɛn ka ɲɛsin bɔli ma minnu bɛ bɔ, sigida minnu bɛ se ka jateminɛ kɛcogo juguw jira hali ni ja kelen-kelen bɛɛ bɛ i n’a fɔ u bɛ cogo la sanfɛ.

Jɛkulu minnu bɛ baarakɛcogo gɛlɛnw ɲɛnabɔ AI yiriwali, fɛn dilannenw seginni, ani jogoɲumanya sabatili siratigɛ la, nin ko suguw b’a jira ko mago bɛ baarakɛcogo yecogo cɛmancɛ la. Ni bug dɔ ye sɛbɛnni kode, waatibolodacogo ani dekode (decoder) lamini, ka sɛgɛsɛgɛli nɔfɛtaama baarakɛminɛn jɛnsɛnnenw ni kumaɲɔgɔnya siraw cɛ minnu tigɛlen don, o bɛ a yɛrɛ ka ɲɔgɔndan kɛ.

Ɲininkali minnu bɛ kɛ tuma caman na

Yala Stable Diffusion pusi debugging ko kɛra ko ye min kɛra tiɲɛ na wa ?

Maana jɔnjɔn sinsinnen bɛ jatebɔ dɔ kan min tilalen don kosɛbɛ ka bɔ AI ɛntɛrinɛti jɛkulu fɛ san 2023. Hali n’a y’a sɔrɔ ko kunnafoni kɛrɛnkɛrɛnnenw kɛra nsiirin ye dɔɔnin lakali seginni na, fɛɛrɛko siratigɛ min bɛ a kɔnɔ — klaviyeti doncogo k’a sɔrɔ a ma kɛ cogo si la min bɛ yɔrɔ dogolen dɔ bɔ kɛnɛ kan — o sɛbɛnnen don koɲuman, wa a bɛ bɛn ni dɛsɛ cogoya dɔntaw ye diffusion model architectures kɔnɔ. O ɲɔgɔnna sɔrɔli kasaaralenw kɛra porozɛw dilanni tariku bɛɛ kɔnɔ.

yala fuzz testing bɛ se ka buguriw minɛ ni dannaya ye AI modeliw kɔnɔ minnu bɛ bɔ mɔgɔw la wa ?

Fuzz sɛgɛsɛgɛli bɛ se ka kɛ ka buguri suguya dɔw minɛ, kɛrɛnkɛrɛnnenya la minnu ɲɛsinnen bɛ donta parsing ma, tokenization edge cases, ani jatebɔ sabatili koɲɛw. Nka, a tɛ warijɛ marifa ye AI sɛnɛcogo la. Ikomi nin misali ninnu bɛ bɔli kɛ dantigɛliw la sanni ka kɛ dantigɛliw ye, min bɛ kɛ "dɛsɛ" ye fuzz sɛgɛsɛgɛli waati la, o ɲɛfɔli bɛ anomali dɔnni fɛɛrɛ caman de wajibiya sanni ka tɛmɛsira/dɛsɛko nɔgɔmanw de wajibiya.

AI baarakɛlaw ka jɛkuluw bɛ baarakɛcogo debugging ɲɛnabɔ cogo di sistɛmu gɛlɛnw kɔnɔ ?

AI jɛkulu kɔgɔlen fanba bɛ u jigi da kɔrɔbɔli nɔfɛtaamayɔrɔw kan, jirisunw tigɛcogo cɛmancɛ la, jɛkafɔ sɛbɛnw, ani porozɛw ɲɛnabɔli sigilenw faralen ɲɔgɔn kan. Gɛlɛyaba min bɛ yen, o ye ka tugu-tugu-ɲɔgɔn kɔ — ka bɔlifɛn kɛrɛnkɛrɛnnen dɔ siri modeli version (modèle version) la, kalan kunnafoniw, hyperparamètres, ani code commit min y’a dilan. Ekipu minnu bɛ o baarakɛcogo ninnu fara ɲɔgɔn kan ka kɛ baarakɛcogo kelenw ye, olu bɛ waati dɔgɔya kosɛbɛ jɛkafɔ kunbabaw la, ka waati caman kɛ gɛlɛyaw ɲɛnabɔli yɛrɛ la.

I ka baarakɛcogo gɛlɛya nɔgɔya

I mana AI modɛliw debugging kɛ walima ka jagokɛcogo gɛlɛn wɛrɛw ɲɛnabɔ, baarakɛminɛn tilalenw bɛ miiri tilalenw dilan. Mewayz bɛ na ni modulu 207 ye minnu bɛ ɲɔgɔn kan, ka kɛ jagokɛyɔrɔ kelen ye, baarakɛla 138.000 ni kɔ dalen bɛ min na — ka yecogo cɛmancɛ di i ka kulu ma, min ka kan ka gɛlɛyaw nɔfɛtaama u bɔyɔrɔ la, ka jaabiw labɛn, ani ka taa teliya la. aw ka kɔrɔbɔli fu daminɛ app.mewayz.com kan k' a lajɛ baarakɛcogo kelenw bɛ cogo min na .