Ntau lub ntiaj teb no tau pib nqis peev ntau hauv kev kawm tshuab (ML).
ML qauv tuaj yeem pib pib thiab ua haujlwm los ntawm pab pawg ntawm cov kws tshaj lij, tab sis ib qho ntawm cov teeb meem loj tshaj plaws yog hloov cov kev paub tau txais mus rau cov qauv tom ntej kom cov txheej txheem nthuav dav.
Txhawm rau txhim kho thiab tsim cov txheej txheem koom nrog hauv cov qauv kev tswj hwm lub neej, cov txheej txheem MLOps tau siv ntau dua los ntawm cov pab pawg uas tsim cov qauv kev kawm tshuab.
Nyeem ntxiv kom paub ntau ntxiv txog qee qhov zoo tshaj plaws MLOps cov cuab yeej thiab cov platforms muaj niaj hnub no thiab lawv tuaj yeem ua kom lub tshuab kawm yooj yim dua los ntawm lub cuab yeej, tus tsim tawm, thiab cov txheej txheem sawv cev.
MLOps yog dab tsi?
Ib txoj hauv kev los tsim cov cai, cov qauv, thiab cov kev coj ua zoo tshaj plaws rau cov qauv kev kawm tshuab hu ua "machine learning operations," lossis "MLOps."
MLOps lub hom phiaj los lav tag nrho lub neej voj voog ntawm ML txoj kev loj hlob - los ntawm kev xav mus rau kev xa tawm - yog cov ntaub ntawv zoo thiab tswj kom tau txais txiaj ntsig zoo tshaj plaws tsis yog nqis peev ntau lub sijhawm thiab cov peev txheej hauv nws yam tsis muaj lub tswv yim.
Lub hom phiaj ntawm MLOps yog txhawm rau txhim kho cov kev coj ua zoo tshaj plaws hauv txoj hauv kev uas ua rau kev txhim kho tshuab kev kawm ntau dua rau ML cov tswv lag luam thiab cov tsim tawm, nrog rau txhim kho qhov zoo thiab kev ruaj ntseg ntawm ML qauv.
Qee tus xa mus rau MLOps li "DevOps rau kev kawm tshuab" txij li nws tau ua tiav siv DevOps cov hauv paus ntsiab lus rau qhov chaw tshwj xeeb ntawm kev txhim kho thev naus laus zis.
Qhov no yog ib txoj hauv kev zoo los xav txog MLOps vim tias, zoo li DevOps, nws hais txog kev sib koom kev paub, kev sib koom tes, thiab kev coj ua zoo tshaj plaws ntawm pab pawg thiab cov cuab yeej.
MLOps muab cov neeg tsim khoom, cov kws tshawb fawb cov ntaub ntawv, thiab pab pawg ua haujlwm nrog lub hauv paus rau kev koom tes thiab, vim li ntawd, tsim cov qauv ML muaj zog tshaj plaws.
Vim li cas thiaj siv cov cuab yeej MLOps?
MLOps cov cuab yeej tuaj yeem ua tau ntau yam dej num rau pab pawg ML, txawm li cas los xij, lawv feem ntau faib ua ob pawg: kev tswj hwm platform thiab kev tswj hwm tus kheej.
Thaum qee cov khoom lag luam MLOps tsuas yog tsom mus rau ib qho tseem ceeb ua haujlwm, xws li cov ntaub ntawv lossis kev tswj hwm metadata, lwm yam cuab yeej siv cov tswv yim ntau dua thiab muab MLOps platform los tswj ntau yam ntawm ML lifecycle.
Saib rau MLOps cov kev daws teeb meem uas pab koj pab neeg hauv kev tswj hwm cov chaw tsim kho ML no, txawm tias koj tab tom nrhiav tus kws tshaj lij lossis cov cuab yeej dav dua:
- Kev tswj cov ntaub ntawv
- Tsim thiab qauv
- Kev tswj cov haujlwm thiab chaw ua haujlwm
- ML qauv xa tawm thiab kev tu ncua tsis tu ncua
- Kev tswj hwm lub neej txij thaum pib mus txog thaum xaus, uas feem ntau yog muab los ntawm MLOps cov kev pabcuam puv ntoob.
MLOps Tools
1. MLFlow
Lub tshuab kev kawm lub neej yog tswj los ntawm qhov qhib-qhov platform MLflow thiab suav nrog cov qauv hauv nruab nrab sau npe, xa mus, thiab kev sim.
MLflow tuaj yeem siv los ntawm ib pab pawg loj, ob leeg ib leeg thiab koom ua ke. Cov tsev qiv ntawv tsis muaj kev cuam tshuam rau lub cuab yeej.
Txhua yam lus programming thiab lub tsev qiv ntawv kawm tshuab tuaj yeem siv tau.
Txhawm rau ua kom yooj yim rau kev cob qhia, xa mus, thiab tswj kev siv tshuab kev kawm, MLFlow cuam tshuam nrog ntau lub tshuab kev kawm, suav nrog TensorFlow thiab Pytorch.
Tsis tas li ntawd, MLflow muab APIs yooj yim-rau-siv uas tuaj yeem suav nrog hauv ib qho kev kawm tshuab lossis cov tsev qiv ntawv uas twb muaj lawm.
MLflow muaj plaub yam tseem ceeb uas pab txhawb kev taug qab thiab npaj kev sim:
- MLflow Taug qab - API thiab UI rau kev nkag mus rau hauv tshuab kev kawm code tsis, versions, metrics, thiab artifacts nrog rau tom qab tso tawm thiab sib piv cov txiaj ntsig
- MLflow Projects – ntim tshuab kawm code nyob rau hauv ib tug reusable, reproducible hom ntawv rau kev hloov mus rau ntau lawm los yog sib koom nrog lwm cov ntaub ntawv zaum
- MLflow Models - tswj thiab xa cov qauv mus rau ntau yam qauv kev pabcuam thiab kev xav tau los ntawm ntau lub tsev qiv ntawv ML
- MLflow Model Registry - lub hauv paus qauv khw uas ua rau muaj kev koom tes tswj hwm tus qauv MLflow tag nrho lub neej, suav nrog cov qauv versioning, theem hloov, thiab cov lus piav qhia.
2. KubeFlow
ML toolbox rau Kubernetes hu ua Kubeflow. Ntim thiab tswj Docker ntim, pab hauv kev saib xyuas ntawm tshuab kev kawm tshuab.
Los ntawm simplifying khiav orchestration thiab deployments ntawm machine learning workflows, nws txhawb lub scalability ntawm tshuab kawm qauv.
Nws yog qhov qhib qhov project uas suav nrog ib pab pawg ua tib zoo xaiv cov cuab yeej ntxiv thiab cov qauv tsim kom haum rau cov kev xav tau ML sib txawv.
Kev cob qhia ML ntev, kev sim phau ntawv, rov ua dua, thiab DevOps cov nyom tuaj yeem daws nrog Kubeflow Pipelines.
Rau ntau theem ntawm kev kawm tshuab, suav nrog kev cob qhia, kev tsim cov kav dej, thiab kev saib xyuas ntawm Jupyter phau ntawv, Kubeflow muaj cov kev pabcuam tshwj xeeb thiab kev koom ua ke.
Nws ua rau nws yooj yim los tswj thiab taug qab lub neej ntawm koj cov haujlwm AI nrog rau kev siv tshuab kev kawm (ML) qauv thiab cov ntaub ntawv xa mus rau Kubernetes pawg.
Nws muaj:
- Cov phau ntawv rau kev siv SDK los cuam tshuam nrog lub kaw lus
- tus neeg siv interface (UI) rau kev tswj thiab saib xyuas kev khiav haujlwm, kev ua haujlwm, thiab kev sim
- Txhawm rau tsim cov kev daws teeb meem kawg-rau-kawg sai yam tsis tas yuav rov tsim dua txhua lub sijhawm, thiab rov siv cov khoom siv thiab cov kav dej.
- Raws li ib feem tseem ceeb ntawm Kubeflow lossis raws li kev teeb tsa ib leeg, Kubeflow Pipelines muaj.
3. Data Version Control
Ib qho qhib-qhov kev tswj hwm kev daws teeb meem rau cov phiaj xwm kev kawm tshuab hu ua DVC, lossis Cov Ntaub Ntawv Version Tswj.
Txawm koj xaiv hom lus twg los xij, nws yog ib qho kev sim uas pab hauv kev txhais cov kav dej.
DVC siv cov cai, cov ntaub ntawv versioning, thiab rov tsim dua tshiab los pab koj txuag lub sijhawm thaum koj pom qhov teeb meem nrog tus qauv ua ntej ntawm koj tus qauv ML.
Tsis tas li ntawd, koj tuaj yeem siv DVC pipelines los cob qhia koj tus qauv thiab faib rau koj cov neeg koom tes. Lub koom haum loj cov ntaub ntawv thiab versioning tuaj yeem ua haujlwm los ntawm DVC, thiab cov ntaub ntawv tuaj yeem khaws cia hauv qhov yooj yim nkag mus.
Txawm hais tias nws suav nrog qee qhov (tseem txwv) kev sim taug qab cov yam ntxwv, nws feem ntau tsom rau cov ntaub ntawv thiab cov raj xa dej thiab kev tswj hwm.
Nws muaj:
- Nws yog cia agnostic, yog li nws muaj peev xwm ua hauj lwm ntau hom cia.
- Nws muab kev taug qab stats ib yam nkaus.
- ib qho kev tsim ua ntej ntawm kev koom nrog ML theem mus rau hauv DAG thiab khiav tag nrho cov kav dej txij thaum pib mus rau qhov kawg
- Txhua tus qauv ML txoj kev txhim kho tag nrho tuaj yeem ua raws li siv nws cov cai thiab cov ntaub ntawv pov thawj.
- Reproducibility los ntawm kev ncaj ncees khaws cia qhov pib teeb tsa, cov ntaub ntawv tawm tswv yim, thiab cov cai hauv kev ua haujlwm rau kev sim.
4. Pachyderm
Pachyderm yog qhov kev tswj hwm version rau kev kawm tshuab thiab cov ntaub ntawv tshawb fawb, zoo ib yam li DVC.
Tsis tas li ntawd, vim nws tau tsim siv Docker thiab Kubernetes, nws tuaj yeem ua tiav thiab siv Machine Learning daim ntawv thov ntawm txhua lub platform huab.
Pachyderm ua kom lav tias txhua daim ntaub ntawv uas tau noj rau hauv cov qauv kev kawm tshuab tuaj yeem taug qab rov qab thiab hloov kho.
Nws yog siv los tsim, faib, tswj, thiab khaws lub qhov muag ntawm cov qauv kev kawm tshuab. Daim ntawv teev npe qauv, tus qauv kev tswj hwm, thiab CLI toolbox tag nrho suav nrog.
Cov neeg tsim khoom tuaj yeem siv tau thiab nthuav dav lawv lub tshuab kev kawm lub neej siv Pachyderm cov ntaub ntawv hauv paus, uas tseem ua kom rov ua dua.
Nws txhawb nqa cov txheej txheem tswj hwm cov ntaub ntawv nruj, txo cov ntaub ntawv ua haujlwm thiab cov nqi khaws cia, thiab pab cov lag luam coj lawv cov ntaub ntawv tshawb fawb pib ua lag luam sai dua.
5. Polyaxon
Siv Polyaxon platform, kev kawm tshuab thiab cov ntawv thov kev kawm sib sib zog nqus tuaj yeem rov ua dua thiab tswj hwm lawv lub neej tag nrho.
Polyaxon muaj peev xwm los tuav thiab tswj cov cuab yeej, thiab nws tuaj yeem muab tso rau hauv txhua qhov chaw khaws ntaub ntawv lossis chaw muab huab. Xws li Torch, Tensorflow, thiab MXNet, uas txhawb nqa tag nrho cov kev kawm sib sib zog nqus nrov tshaj plaws.
Thaum nws los txog rau orchestration, Polyaxon ua rau koj ua tau ntau tshaj ntawm koj pawg los ntawm kev teem sijhawm ua haujlwm thiab kev xeem ntawm lawv CLI, dashboard, SDKs, lossis REST API.
Nws muaj:
- Koj tuaj yeem siv qhov qhib-qhov version tam sim no, tab sis nws kuj suav nrog kev xaiv rau cov tuam txhab.
- Txawm hais tias nws npog tag nrho lub neej voj voog, suav nrog kev khiav orchestration, nws muaj peev xwm ntau ntxiv.
- Nrog rau kev siv cov ntaub ntawv, pib cov lus qhia, cov ntaub ntawv kawm, phau ntawv qhia, kev qhia, kev hloov pauv, thiab lwm yam, nws yog lub platform uas muaj ntaub ntawv zoo heev.
- Nrog rau qhov kev soj ntsuam kev pom dashboard, nws muaj peev xwm ua kom qhov muag ntawm, taug qab, thiab ntsuas txhua qhov kev sim ua kom zoo dua.
6. Comet
Comet yog lub platform rau kev kawm tshuab meta uas taug qab, sib piv, piav qhia, thiab txhim kho kev sim thiab cov qauv.
Tag nrho koj qhov kev sim tuaj yeem pom thiab muab piv rau hauv ib qho chaw.
Nws ua haujlwm rau txhua txoj haujlwm kev kawm tshuab, nyob qhov twg koj cov lej ua tiav, thiab nrog rau txhua lub tsev qiv ntawv kawm tshuab.
Comet yog tsim nyog rau cov pab pawg, cov tib neeg, cov tsev kawm ntawv, cov lag luam, thiab lwm tus neeg uas xav kom pom kev sim sai sai, ua kom yooj yim rau kev ua haujlwm, thiab ua qhov kev sim.
Cov kws tshawb fawb cov ntaub ntawv thiab pab pawg tuaj yeem taug qab, qhia meej, txhim kho, thiab sib piv cov kev sim thiab cov qauv siv tus kheej-hosted thiab huab-based meta-machine kev kawm platform Comet.
Nws muaj:
- Muaj ntau lub peev xwm muaj rau cov neeg koom tes sib koom ua haujlwm.
- Nws muaj ntau qhov kev sib koom ua ke uas ua rau nws yooj yim los txuas nws mus rau lwm yam thev naus laus zis
- Ua haujlwm zoo nrog cov tsev qiv ntawv ML tam sim no
- Saib xyuas cov neeg siv kev tswj hwm
- Kev sib piv ntawm cov kev sim tau qhib, suav nrog kev sib piv ntawm cov lej, hyperparameters, metrics, kev kwv yees, kev vam khom, thiab kev ntsuas qhov system.
- Muab cov qauv sib txawv rau kev pom, suab, ntawv nyeem, thiab cov ntaub ntawv tabular uas cia koj pom cov qauv.
7. Optuna
Optuna yog ib qho system rau autonomous hyperparameter optimization uas tuaj yeem siv rau ob qho tib si kev kawm tshuab thiab kev kawm tob nrog rau lwm yam.
Nws muaj ntau yam kev txiav-ntug algorithms uas koj tuaj yeem xaiv (lossis txuas), ua rau nws yooj yim heev los faib kev cob qhia ntau lub khoos phis tawj, thiab muab cov txiaj ntsig zoo nkauj pom.
Cov tsev qiv ntawv nrov tshuab kawm xws li PyTorch, TensorFlow, Keras, FastAI, sci-kit-kawm, LightGBM, thiab XGBoost yog tag nrho nrog nws.
Nws muab cov txheej txheem txiav-ntug uas ua rau cov neeg siv khoom tau txais txiaj ntsig sai dua los ntawm kev txo qis cov qauv uas tsis zoo li pheej hmoo.
Siv Python-based algorithms, nws cia li tshawb nrhiav qhov zoo tshaj plaws hyperparameters. Optuna txhawb kom parallelized hyperparameter tshawb nrhiav thoob plaws ntau cov xov tsis hloov pauv tus lej qub.
Nws muaj:
- Nws txhawb nqa kev cob qhia faib rau ib pawg nrog rau ib lub computer (multi-process) (multi-node)
- Nws txhawb nqa ob peb txoj kev trimming kom ceev convergence (thiab siv tsawg xam)
- Nws muaj ntau yam kev pom muaj zog, xws li daim duab kos duab, cov duab kos duab, thiab cov kab lus sib dhos.
8. Kedro
Kedro yog Python lub moj khaum pub dawb rau kev sau cov lej uas tuaj yeem hloov kho thiab khaws cia rau cov ntaub ntawv tshawb fawb.
Nws coj cov tswv yim los ntawm kev coj ua zoo tshaj plaws hauv software engineering rau tshuab kev kawm code. Python yog lub hauv paus ntawm qhov kev ua haujlwm orchestration cuab yeej.
Txhawm rau ua kom koj cov txheej txheem ML yooj yim dua thiab meej dua, koj tuaj yeem tsim kho dua tshiab, tswj tau, thiab ua haujlwm ua haujlwm.
Kedro suav nrog software engineering cov hauv paus ntsiab lus xws li kev hloov pauv, kev sib cais ntawm lub luag haujlwm, thiab hloov pauv mus rau hauv ib puag ncig kev kawm tshuab.
Raws li lub hauv paus ntawm Cookiecutter Data Science, nws muab ib qho kev sib koom ua ke, hloov kho qhov project.
Ib tug xov tooj ntawm cov ntaub ntawv yooj yim connectors siv los khaws thiab thauj cov ntaub ntawv hla ntau cov ntaub ntawv systems thiab cov ntaub ntawv tawm tswv yim, yog tswj los ntawm cov ntaub ntawv catalog. Nws ua rau kev kawm tshuab ua haujlwm zoo dua thiab ua rau nws yooj yim dua los tsim cov ntaub ntawv kav dej.
Nws muaj:
- Kedro tso cai rau dispersed los yog tib lub tshuab xa mus.
- Koj tuaj yeem automate dependencies ntawm Python code thiab workflow visualization siv pipeline abstraction.
- Los ntawm kev siv modular, reusable code, cov cuab yeej no pab txhawb pab pawg sib koom ua ke ntawm ntau qib thiab txhim kho cov khoom tsim tau hauv ib puag ncig coding.
- Lub hom phiaj tseem ceeb yog txhawm rau kov yeej qhov tsis zoo ntawm Jupyter phau ntawv sau, cov ntawv sau ib zaug, thiab cov kua nplaum los ntawm kev sau cov ntaub ntawv khaws cia kev tshawb fawb.
9. BentoML
Lub tsev kawm tshuab kawm API qhov kawg tau ua kom yooj yim dua nrog BentoML.
Nws muab ib qho tseem ceeb tsis tau condensed infrastructure txav mus kawm tshuab kev kawm qauv rau hauv ntau lawm.
Nws tso cai rau koj los ntim cov qauv kawm rau kev siv hauv qhov chaw tsim khoom, txhais lawv siv ib qho kev ua haujlwm ML. Ob qho offline batch pabcuam thiab online API pabcuam tau txais kev txhawb nqa.
Ib tus qauv ua haujlwm siab thiab ua haujlwm yooj yim yog cov yam ntxwv ntawm BentoML.
Tsis tas li ntawd, tus neeg rau zaub mov muaj kev hloov kho micro-batch. Ib qho kev sib koom ua ke rau kev teeb tsa cov qauv thiab ua raws cov txheej txheem xa tawm yog muab los ntawm UI dashboard.
Yuav tsis muaj neeg rau zaub mov downtime vim hais tias kev khiav hauj lwm mechanism yog modular thiab configuration yog reusable. Nws yog lub platform hloov tau yooj yim rau kev muab, teeb tsa, thiab siv ML qauv.
Nws muaj:
- Nws muaj ib tug modular tsim uas yog adaptable.
- Nws tso cai rau kev xa tawm thoob plaws ntau lub platform.
- Nws tsis tuaj yeem txiav qhov kev ntsuas kab rov tav.
- Nws tso cai rau ib qho qauv qauv, kev tswj tus qauv, qauv ntim, thiab kev ua haujlwm siab ua haujlwm.
10. seldon
Cov kws tshawb fawb cov ntaub ntawv tuaj yeem tsim, siv, thiab tswj cov qauv kev kawm tshuab thiab kev sim ntawm qhov ntsuas ntawm Kubernetes siv qhov qhib-qhov Seldon Core moj khaum.
TensorFlow, sci-kit-kawm, Spark, R, Java, thiab H2O tsuas yog ob peb yam khoom siv uas tau txais kev txhawb nqa los ntawm nws.
Nws tseem cuam tshuam nrog Kubeflow thiab RedHat's OpenShift. Cov tub ntxhais Seldon hloov cov qauv kev kawm tshuab (ML qauv) lossis cov ntaub ntawv hais lus (lus zoo li Python, Java, thiab lwm yam) rau hauv kev tsim khoom REST/GRPC microservices.
Ib qho zoo tshaj plaws MLOps cov cuab yeej rau kev txhim kho cov txheej txheem kev kawm tshuab yog qhov no.
Nws yog qhov yooj yim rau ntim cov qauv ML thiab ntsuas kev siv tau thiab kev nyab xeeb siv Seldon Core.
Nws muaj:
- Kev xa tawm qauv tuaj yeem ua kom yooj yim nrog ntau txoj hauv kev, xws li kev xa tawm canary.
- Txhawm rau kom nkag siab tias vim li cas qhov kev kwv yees tshwj xeeb raug tsim, siv cov qauv piav qhia.
- Thaum muaj teeb meem tshwm sim, khaws lub qhov muag ntawm cov qauv tsim khoom siv qhov kev ceeb toom.
xaus
MLOps tuaj yeem pab ua kom lub tshuab kawm ua haujlwm zoo dua. MLOps tuaj yeem ua kom muaj kev xa tawm sai, ua kom cov ntaub ntawv khaws cia thiab kev debugging yooj yim dua, thiab txhim kho kev sib koom tes ntawm cov kws tshaj lij thiab cov kws tshawb fawb cov ntaub ntawv.
Txhawm rau kom koj xaiv MLOps cov cuab yeej uas zoo tshaj plaws rau koj cov kev xav tau, cov ntawv tshaj tawm no tau tshuaj xyuas 10 cov kev daws teeb meem nrov MLOps, feem ntau yog qhib qhov chaw.
Sau ntawv cia Ncua