Ngo-2018, umqondo wama-MLOps wavela emibuthanweni yobungcweti nasezingqungqutheleni ezinezihloko ezinikezelwe ku-AI, ezabamba ngokushesha embonini futhi manje ezithuthukayo njengesiqondiso esizimele. Ngokuzayo, ama-MLOps angase abe enye yezindawo ezidume kakhulu ku-IT. Kuyini futhi kudliwa ngani? Ake sithole ngezansi.
Yini i-MLOps
Ama-MLOps (okuhlanganisa ubuchwepheshe bokufunda komshini nezinqubo nezindlela zokuqalisa amamodeli athuthukisiwe kuzinqubo zebhizinisi) ayindlela entsha yokusebenzisana phakathi kwabamele ibhizinisi, ososayensi, ochwepheshe bezibalo, ochwepheshe bokufunda ngomshini nonjiniyela be-IT lapho bedala amasistimu obuhlakani bokwenziwa.
Ngamanye amazwi, kuyindlela yokuguqula izindlela zokufunda ngomshini nobuchwepheshe kube ithuluzi eliwusizo lokuxazulula izinkinga zebhizinisi.
Kubalulekile ukuqonda ukuthi uchungechunge lokukhiqiza luqala isikhathi eside ngaphambi kokuthuthukiswa kwemodeli. Isinyathelo sayo sokuqala ukuchaza inkinga yebhizinisi, inkoleloze mayelana nevelu engakhishwa kudatha, kanye nombono webhizinisi wokuwusebenzisa.
Wona kanye umqondo we-MLOps uqhamuke njengesifaniso somqondo we-DevOps maqondana namamodeli okufunda komshini nobuchwepheshe. I-DevOps iyindlela yokuthuthukisa isofthiwe ekuvumela ukuthi ukhuphule isivinini sokuqaliswa kwezinguquko zomuntu ngamunye ngenkathi ugcina ukuguquguquka nokuthembeka usebenzisa izindlela eziningi, okuhlanganisa ukuthuthukiswa okuqhubekayo, ukuhlukaniswa kwemisebenzi ibe yinani lama-microservices azimele, ukuhlola okuzenzakalelayo kanye nokusetshenziswa komuntu ngamunye. izinguquko, ukuqapha kwezempilo emhlabeni wonke, uhlelo lokuphendula ngokushesha ngokwehluleka okutholiwe, njll.
I-DevOps ichaze umjikelezo wempilo wesoftware, futhi umphakathi uqhamuke nombono wokusebenzisa indlela efanayo kudatha enkulu. I-DataOps ingumzamo wokujwayela nokwandisa indlela yokusebenza kucatshangelwa izici zokugcina, ukudlulisa nokucubungula inani elikhulu ledatha ezisekelweni ezihlukahlukene nezisebenzisanayo.
Ngokufika kwenqwaba ethile ebucayi yamamodeli okufunda omshini asetshenziswe ezinqubweni zebhizinisi zezinkampani, ukufana okuqinile kuye kwaqashelwa phakathi komjikelezo wempilo wamamodeli okufunda omshini wezibalo kanye nomjikelezo wempilo wesofthiwe. Umehluko kuphela ukuthi ama-algorithms emodeli adalwa kusetshenziswa amathuluzi okufunda omshini nezindlela. Ngakho-ke, umqondo wavela ngokwemvelo wokusebenzisa nokuvumelanisa izindlela ezaziwayo kakade ekuthuthukisweni kwesofthiwe yamamodeli okufunda omshini. Ngakho-ke, izigaba ezibalulekile ezilandelayo zingahlukaniswa emjikelezweni wokuphila wamamodeli wokufunda womshini:
- ukuchaza umqondo webhizinisi;
- ukuqeqeshwa okuyimodeli;
- ukuhlola nokusebenzisa imodeli enqubweni yebhizinisi;
- ukusebenza kwemodeli.
Uma ngesikhathi sokusebenza kunesidingo sokushintsha noma ukuqeqesha kabusha imodeli kudatha entsha, umjikelezo uqala futhi - imodeli iyacwengwa, ihlolwe, futhi inguqulo entsha iyasetshenziswa.
Hlehla. Kungani uziqeqeshe futhi ungaphindi uziqeqeshe? Igama elithi "ukuqeqeshwa kabusha kwemodeli" linencazelo ekabili: phakathi kochwepheshe lisho iphutha lemodeli, lapho imodeli ibikezela kahle, empeleni iphinda ipharamitha ebikezelwe kusethi yokuqeqesha, kodwa yenza okubi kakhulu kusampula yedatha yangaphandle. Ngokwemvelo, imodeli enjalo iyisici, ngoba lesi sici asikuvumeli ukusetshenziswa kwayo.
Kulo mjikelezo wempilo, kubonakala kunengqondo ukusebenzisa amathuluzi e-DevOps: ukuhlola okuzenzakalelayo, ukuthunyelwa nokuqapha, ukuklama izibalo zamamodeli ngendlela yama-microservices ahlukene. Kodwa futhi kukhona izici eziningi ezivimbela ukusetshenziswa okuqondile kwalawa mathuluzi ngaphandle kokubophezela kwe-ML eyengeziwe.
Indlela yokwenza amamodeli asebenze futhi abe nenzuzo
Njengesibonelo lapho sizobonisa khona ukusetshenziswa kwendlela ye-MLOps, sizothatha umsebenzi wakudala wokwenza amarobhothi usekelo lwengxoxo lomkhiqizo wasebhange (noma yimuphi omunye). Ngokuvamile, inqubo yebhizinisi yokusekela ingxoxo ibukeka kanje: iklayenti lifaka umlayezo onombuzo engxoxweni futhi lithole impendulo evela kuchwepheshe ongaphakathi kwesihlahla sengxoxo esichazwe ngaphambilini. Umsebenzi wokuzenzakalela ingxoxo enjalo ngokuvamile uxazululwa kusetshenziswa amasethi emithetho echazwe ngochwepheshe, edinga umsebenzi onzima kakhulu ukuyithuthukisa nokuyigcina. Ukusebenza kahle kwe-automation enjalo, kuye ngezinga lobunzima bomsebenzi, kungaba ngu-20-30%. Ngokwemvelo, kuphakama umbono wokuthi kunenzuzo enkulu ukwethula imojula yobuhlakani bokufakelwa - imodeli eyenziwe kusetshenziswa umshini wokufunda, okuthi:
- iyakwazi ukucubungula inani elikhulu lezicelo ngaphandle kokubamba iqhaza komsebenzisi (kuye ngokuthi isihloko, kwezinye izimo ukusebenza kahle kungafinyelela ku-70-80%);
- ijwayelana kangcono namagama angajwayelekile engxoxweni - iyakwazi ukunquma inhloso, isifiso sangempela somsebenzisi ngokusekelwe esicelweni esingakhiwanga ngokucacile;
- uyazi ukuthi anganquma kanjani ukuthi impendulo yemodeli yanele nini, futhi uma kukhona ukungabaza mayelana "nokuqwashisa" kwalesi mpendulo futhi udinga ukubuza umbuzo owengeziwe wokucacisa noma ushintshele ku-opharetha;
- ingaqeqeshwa ngokungeziwe ngokuzenzakalelayo (esikhundleni sokuthi iqembu lonjiniyela lihlale lizivumelanisa nezimo futhi lilungisa izikripthi zokuphendula, imodeli ibuye iqeqeshwe uchwepheshe Wesayensi Yedatha isebenzisa imitapo yolwazi yokufunda yomshini efanele).
Indlela yokwenza imodeli enjalo ethuthukisiwe isebenze?
Njengasekuxazululeni noma iyiphi enye inkinga, ngaphambi kokwenza imojuli enjalo, kuyadingeka ukuchaza inqubo yebhizinisi futhi uchaze ngokusemthethweni umsebenzi othile esizowuxazulula sisebenzisa indlela yokufunda yomshini. Kuleli qophelo, inqubo yokusebenza, eqokwe isifinyezo se-Ops, iyaqala.
Isinyathelo esilandelayo ukuthi i-Data Scientist, ngokubambisana Nonjiniyela Wedatha, ihlola ukutholakala nokwanela kwedatha kanye ne-hypothesis yebhizinisi mayelana nokusebenza kombono webhizinisi, ukuthuthukisa imodeli yesibonelo kanye nokuhlola ukusebenza kwayo kwangempela. Kungemva kokuqinisekiswa yibhizinisi kuphela lapho uguquko lungaqala khona ekuthuthukiseni imodeli ukuya ekulihlanganiseni nezinhlelo ezenza inqubo ethile yebhizinisi. Ukuhlelwa kokuqaliswa kokuphela kuze kube sekupheleni, ukuqonda okujulile esigabeni ngasinye sokuthi imodeli izosetshenziswa kanjani nokuthi izoletha muphi umphumela wezomnotho, kuyiphuzu elibalulekile ezinqubweni zokwethula izindlela ze-MLOps endaweni yezobuchwepheshe yenkampani.
Ngokuthuthukiswa kobuchwepheshe be-AI, inombolo nenhlobonhlobo yezinkinga ezingaxazululwa kusetshenziswa ukufunda ngomshini zikhula ngamandla. Inqubo ngayinye yebhizinisi enjalo isindisa inkampani ngenxa yokuzenzakalelayo komsebenzi wabasebenzi abaningi (isikhungo socingo, ukuhlola nokuhlunga imibhalo, njll.), kungukunwetshwa kwesisekelo samakhasimende ngokungeza imisebenzi emisha ekhangayo futhi elula, it. yonga imali ngenxa yokusebenzisa kahle nokwabiwa kabusha kwezinsiza nokunye okuningi. Ekugcineni, noma iyiphi inqubo igxile ekudaleni inani futhi, ngenxa yalokho, kufanele ilethe umphumela othile wezomnotho. Lapha kubaluleke kakhulu ukwakha umqondo webhizinisi ngokucacile futhi ubale inzuzo elindelekile ekusebenziseni imodeli kusakhiwo sokudala inani lenkampani. Kunezimo lapho ukusebenzisa imodeli kungazilungisi, futhi isikhathi esichithwa ochwepheshe bokufunda ngomshini sibiza kakhulu kunendawo yokusebenza yomsebenzisi owenza lo msebenzi. Kungakho kudingekile ukuzama ukukhomba amacala anjalo ezigabeni zokuqala zokudala izinhlelo ze-AI.
Ngenxa yalokho, amamodeli aqala ukukhiqiza inzuzo kuphela lapho inkinga yebhizinisi yenziwe ngendlela efanele kunqubo ye-MLOps, izinto ezibalulekile sezibekiwe, futhi inqubo yokwethula imodeli ohlelweni yenziwe ezigabeni zokuqala zentuthuko.
Inqubo entsha - izinselele ezintsha
Impendulo ebanzi yombuzo webhizinisi obalulekile mayelana nokuthi amamodeli e-ML asebenza kanjani ekuxazululeni izinkinga, udaba olujwayelekile lokuthembela ku-AI lungenye yezinselelo ezibalulekile ohlelweni lokuthuthukisa nokusebenzisa izindlela ze-MLOps. Ekuqaleni, amabhizinisi ayangabaza mayelana nokwethulwa kokufundwa komshini ezinqubweni - kunzima ukuthembela kumamodeli ezindaweni lapho ngaphambili, njengomthetho, abantu basebenza khona. Ebhizinisini, izinhlelo zibonakala "ziyibhokisi elimnyama", ukuhambisana kwazo okusadinga ukuqinisekiswa. Ngaphezu kwalokho, emabhange, ebhizinisini labaqhubi be-telecom nabanye, kunezidingo eziqinile zabalawuli bakahulumeni. Wonke amasistimu nama-algorithms asetshenziswa ezinqubweni zebhange angaphansi kokucwaninga. Ukuze kuxazululwe le nkinga, ukufakazela ibhizinisi kanye nabalawuli ukufaneleka nokunemba kwezimpendulo zobuhlakani bokwenziwa, amathuluzi okuqapha ayenziwa kanye nemodeli. Ngaphezu kwalokho, kunenqubo yokuqinisekisa ezimele, okuyimpoqo kumamodeli okulawula, ahlangabezana nezidingo zeBhange Elikhulu. Iqembu lochwepheshe elizimele lihlola imiphumela etholwe imodeli ngokucabangela idatha yokufaka.
Inselele yesibili ukuhlola nokucabangela ubungozi bemodeli lapho usebenzisa imodeli yokufunda yomshini. Ngisho noma umuntu engakwazi ukuphendula umbuzo ngokuqiniseka okungamaphesenti ayikhulu ukuthi leyo ngubo efanayo yayimhlophe noma iluhlaza okwesibhakabhaka, khona-ke ukuhlakanipha okwenziwa nakho kunelungelo lokwenza iphutha. Kuyafaneleka futhi ukucabangela ukuthi idatha ingase ishintshe ngokuhamba kwesikhathi, futhi amamodeli adinga ukuqeqeshwa kabusha ukuze akhiqize umphumela onembe ngokwanele. Ukuqinisekisa ukuthi inqubo yebhizinisi ayihlupheki, kuyadingeka ukuphatha izingozi zemodeli nokuqapha ukusebenza kwemodeli, ukuyiqeqesha njalo kudatha entsha.
Kodwa ngemva kwesigaba sokuqala sokungathembani, umphumela ophambene uqala ukuvela. Lapho amamodeli engeziwe esetshenziswa ngempumelelo ezinqubweni, kulapho umdlandla webhizinisi wokusebenzisa ubuhlakani bokwenziwa ukhula - kutholakala izinkinga ezintsha nezintsha ezingaxazululwa kusetshenziswa izindlela zokufunda ngomshini. Umsebenzi ngamunye uvula inqubo yonke edinga amakhono athile:
- onjiniyela bedatha balungisa futhi bacubungule idatha;
- ososayensi bedatha basebenzisa amathuluzi okufunda omshini futhi bathuthukise imodeli;
- I-IT isebenzisa imodeli ohlelweni;
- Unjiniyela we-ML unquma ukuthi ingahlanganisa kanjani le modeli ngendlela efanele enqubweni, okumele isetshenziswe ngamathuluzi e-IT, kuye ngezidingo zendlela yokusetshenziswa kwemodeli, kucatshangelwa ukugeleza kwezicelo, isikhathi sokuphendula, njll.
- Umakhi we-ML uklama ukuthi umkhiqizo wesofthiwe ungenziwa kanjani ngokoqobo ohlelweni lwezimboni.
Wonke umjikelezo udinga inani elikhulu lochwepheshe abaqeqeshwe kakhulu. Esikhathini esithile ekuthuthukisweni kanye nezinga lokungena kwamamodeli e-ML ezinqubweni zebhizinisi, kuvele ukuthi ukukala inani lochwepheshe ngokulingana nokwanda kwenani lemisebenzi kuyabiza futhi kungasebenzi. Ngakho-ke, umbuzo uvela wokwenza ngokuzenzakalelayo inqubo ye-MLOps - echaza amakilasi ambalwa ajwayelekile ezinkinga zokufunda ngomshini, ukuthuthukisa amapayipi okucubungula idatha ajwayelekile kanye nokuqeqeshwa okwengeziwe kwamamodeli. Esithombeni esifanele, ukuxazulula izinkinga ezinjalo kudinga ochwepheshe abanekhono ngokulinganayo empambanweni ye-Big Data, Data Science, DevOps kanye ne-IT. Ngakho-ke, inkinga enkulu embonini ye-Data Science kanye nenselelo enkulu ekuhleleni izinqubo ze-MLOps ukuntuleka kwamakhono anjalo emakethe yokuqeqesha ekhona. Ochwepheshe abahlangabezana nalezi zidingo okwamanje abayivelakancane emakethe yezabasebenzi futhi bafanele isisindo sabo ngegolide.
Odabeni lwamakhono
Ngokombono, yonke imisebenzi ye-MLOps ingaxazululwa kusetshenziswa amathuluzi e-DevOps yakudala futhi ngaphandle kokusebenzisa isandiso esikhethekile sesibonelo. Khona-ke, njengoba siphawulile ngenhla, usosayensi wedatha akumele abe nje isazi sezibalo kanye nomhlaziyi wedatha, kodwa futhi abe yi-guru yepayipi lonke - unesibopho sokuthuthukisa izakhiwo, amamodeli wokuhlela ngezilimi eziningana kuye ngokuthi izakhiwo, ukulungiselela. i-data mart kanye nokuphakela uhlelo lokusebenza ngokwalo. Kodwa-ke, ukudala uhlaka lobuchwepheshe olusetshenziswa ekuqedeni kuze kube sekugcineni inqubo ye-MLOps kuthatha imali efinyelela ku-80% yezindleko zabasebenzi, okusho ukuthi isazi sezibalo esiqeqeshiwe, esiyikhwalithi yeDatha Scientist, sizonikela kuphela ngama-20% esikhathi saso emsebenzini wakhe okhethekile. . Ngakho-ke, ukuchaza izindima zongcweti ababandakanyekayo ohlelweni lokusebenzisa amamodeli okufunda komshini kubaluleka.
Ukuthi izindima kufanele zichazwe kabanzi kangakanani kuncike kusayizi webhizinisi. Kuyinto eyodwa lapho isiqalisi sinochwepheshe oyedwa, osebenza kanzima endaweni yokugcina amandla, ongunjiniyela wakhe, umakhi wezakhiwo, kanye ne-DevOps. Kuyindaba ehluke ngokuphelele lapho, ebhizinisini elikhulu, zonke izinqubo zokuthuthukisa amamodeli zigxile kochwepheshe abambalwa bezinga eliphezulu be-Data Science, kuyilapho umklami wezinhlelo noma uchwepheshe wedathabhesi - ikhono elivame kakhulu nelingabizi kakhulu emakethe yezabasebenzi - angathatha. emisebenzini eminingi, imisebenzi evamile.
Ngakho-ke, ijubane nekhwalithi yamamodeli athuthukisiwe, ukukhiqiza kweqembu kanye ne-microclimate kulo kuncike ngokuqondile lapho umngcele ulele ekukhethweni kochwepheshe ukusekela inqubo ye-MLOps nokuthi inqubo yokusebenza kwamamodeli athuthukisiwe ihlelwa kanjani. .
Osekwenziwe yiqembu lethu
Muva nje siqale ukwakha uhlaka lwamakhono kanye nezinqubo ze-MLOps. Kodwa amaphrojekthi ethu okuphathwa komjikelezo wempilo kanye nokusebenzisa amamodeli njengesevisi asevele esesigabeni sokuhlola se-MVP.
Siphinde sanquma uhlaka lwamakhono olufanele lwebhizinisi elikhulu kanye nesakhiwo senhlangano sokusebenzisana phakathi kwabo bonke ababambiqhaza kunqubo. Amaqembu e-Agile ahlelwe ukuze axazulule izinkinga kulo lonke uhla lwamakhasimende ebhizinisi, futhi inqubo yokusebenzisana namaqembu ephrojekthi ukudala amapulatifomu nengqalasizinda, okuyisisekelo sesakhiwo se-MLOps esakhiwayo, yasungulwa.
Imibuzo yekusasa
Ama-MLOps ayindawo ekhulayo ebhekene nokushoda kwamakhono futhi izoba nomfutho esikhathini esizayo. Okwamanje, kungcono ukwakha ku-DevOps ukuthuthukiswa kanye nemikhuba. Umgomo oyinhloko wama-MLOps ukusebenzisa ngokuphumelelayo amamodeli e-ML ukuxazulula izinkinga zebhizinisi. Kodwa lokhu kuphakamisa imibuzo eminingi:
- Ungasinciphisa kanjani isikhathi sokwethula amamodeli ekukhiqizeni?
- Unganciphisa kanjani ukungqubuzana kwe-bureaucratic phakathi kwamaqembu wamakhono ahlukene futhi ukwandise ukugxila ekusebenzisaneni?
- Ungawalandela kanjani amamodeli, uphathe izinguqulo futhi uhlele ukuqapha okusebenzayo?
- Ungawakha kanjani umjikelezo wokuphila oyindilinga ngempela wemodeli yesimanje ye-ML?
- Ungalinganisa kanjani inqubo yokufunda komshini?
Izimpendulo zale mibuzo zizocacisa kakhulu ukuthi ama-MLOps azofinyelela amandla awo aphelele ngokushesha kangakanani.
Source: www.habr.com