Ukuqhubela phambili umxholo wokhuphiswano lokufunda koomatshini eHabré, singathanda ukwazisa abafundi kumaqonga amabini angakumbi. Ngokuqinisekileyo azikho zikhulu njenge-kaggle, kodwa ngokuqinisekileyo zifanelwe ukuhoywa.

Ngokomntu, andithandi i-kaggle kakhulu ngenxa yezizathu ezininzi:

• okokuqala, ukhuphiswano luhlala luhlala iinyanga ezininzi, kwaye ukuthatha inxaxheba okusebenzayo kufuna umgudu omkhulu;
• okwesibini, iinkozo zoluntu (izisombululo zoluntu). Abalandeli bakaKaggle bacebisa ukuba babaphathe ngokuzola kweemonki zaseTibet, kodwa eneneni kulihlazo kakhulu xa into obuyisebenzela inyanga okanye ezimbini ngequbuliso iye yabekwa kwisitya sesilivere kumntu wonke.

Ngethamsanqa, ukhuphiswano lokufunda koomatshini lubanjwa kwamanye amaqonga, kwaye isibini solu khuphiswano siya kuxoxwa.

IDAO	SNA Hackathon 2019
Ulwimi olusemthethweni: IsiNgesi, abaququzeleli: Yandex, Sberbank, HSE	Ulwimi olusemthethweni lwesiRashiya, abaququzeleli: Mail.ru Group
Umjikelo we-Intanethi: NgoJanuwari 15 - Feb 11, 2019; Kwindawo yokuGqibela: ku-Epreli 4-6, 2019	kwi-intanethi - ukususela ngoFebruwari 7 ukuya kuMatshi 15; ngaphandle kweintanethi - ukusuka nge-30 kaMatshi ukuya ku-Epreli 1.
Ukusebenzisa iseti ethile yedatha malunga nesuntswana kwi-Large Hadron Collider (i-trajectory, isantya, kunye nezinye iiparameters ezibonakalayo ezinobunzima), misela ukuba yi-muon okanye hayi. Kule nkcazo, imisebenzi emi-2 ichongiwe: - kwenye kufuneka uthumele ingqikelelo yakho, - kwaye kwenye - ikhowudi epheleleyo kunye nemodeli yokuxela kwangaphambili, kwaye ukuphunyezwa kwakuxhomekeke kwizithintelo ezingqongqo ekusebenzeni kwexesha kunye nokusetyenziswa kwememori.	Kukhuphiswano lwe-SNA Hackathon, iilogi zemixholo ebonisa ukusuka kumaqela avulekileyo kwiifidi zeendaba zabasebenzisi ngoFebruwari-Matshi 2018 zaqokelelwa. Isethi yovavanyo iqulethe iveki yokugqibela kunye nesiqingatha sikaMatshi. Ungeno ngalunye kwilogi lunolwazi malunga noko kubonisiwe kwaye kubani, kunye nendlela umsebenzisi aphendule ngayo kulo mxholo: ulinganisile, waphawula, akayihoya, okanye wayifihla kwisondlo. Ingundoqo yemisebenzi ye-SNA ye-Hackathon kuluhlu lomsebenzisi ngamnye womnatha woluntu we-Odnoklassniki ukutya kwakhe, ukuphakamisa phezulu kangangoko kunokwenzeka ezo zithuba eziza kufumana "iklasi". Kwinqanaba le-intanethi, umsebenzi wahlulahlulwe wangamacandelo ama-3: 1. Izikhundla zezikhundla ngokweempawu ezahlukeneyo zentsebenziswano 2. izithuba zenqanaba ngokusekelwe kwimifanekiso equlethwe kuyo 3. hlela izithuba ngokwesicatshulwa ezikuquletheyo
Imetric yesiko elintsonkothileyo, into efana ne-ROC-AUC	I-avareji ye-ROC-AUC ngumsebenzisi
Amabhaso enqanaba lokuqala - iiT-shirts zeendawo eziN, ukudlula ukuya kwinqanaba lesibini, apho indawo yokuhlala kunye nokutya kwahlawulwa ngexesha lokhuphiswano. Isigaba sesibini - ??? (Ngenxa yezizathu ezithile, bendingekho kumsitho wamabhaso kwaye andikwazanga ukufumanisa ukuba athini amabhaso ekugqibeleni). Bathembise ngeelaptops kuwo onke amalungu eqela eliphumeleleyo	Amabhaso kwinqanaba lokuqala - i-T-shirts kubathathi-nxaxheba be-100 abagqwesileyo, ukuhamba ukuya kwinqanaba lesibini, apho uhambo oluya eMoscow, indawo yokuhlala kunye nokutya ngexesha lokhuphiswano lwahlawulwa. Kwakhona, ekupheleni kwenqanaba lokuqala, amabhaso abhengezwa ngeyona nto ibhetele kwimisebenzi emi-3 kwinqanaba loku-1: wonke umntu uphumelele ikhadi levidiyo le-RTX 2080 TI! Inqanaba lesibini yayiyiqela leqela, amaqela aquka abantu aba-2 ukuya kwaba-5, amabhaso: Indawo ye-1 - i-ruble ye-300 Indawo ye-2 - i-ruble ye-200 Indawo ye-3 - i-ruble ye-100 jury ibhaso - 100 ruble
Iqela letelegram esemthethweni, ~ 190 abathathi-nxaxheba, unxibelelwano ngesiNgesi, imibuzo kwafuneka ilinde iintsuku ezininzi ukuze ifumane impendulo.	Iqela elisemthethweni kwitelegram, ~ 1500 abathathi-nxaxheba, ingxoxo esebenzayo yemisebenzi phakathi kwabathathi-nxaxheba kunye nabaququzeleli
Abaququzeleli banikeze izisombululo ezibini ezisisiseko, ezilula neziphambili. Ulula olufunekayo ngaphantsi kwe-16 GB ye-RAM, kwaye imemori eqhubela phambili ayizange ingene kwi-16. Ngexesha elifanayo, ukujonga phambili kancinci, abathathi-nxaxheba abazange bakwazi ukugqithisa kakhulu isisombululo esiphezulu. Akubangakho ngxaki ekuqaliseni ezi zisombululo. Kufuneka kuqatshelwe ukuba kumzekelo ophambili kwakukho inkcazo kunye nenkcazo apho ungaqala khona ukuphucula isisombululo.	Izisombululo zamandulo ezisisiseko zanikezelwa kumsebenzi ngamnye, oye wadlulwa ngokulula ngabathathi-nxaxheba. Kwiintsuku zokuqala zokhuphiswano, abathathi-nxaxheba badibana nobunzima obuninzi: okokuqala, idatha yanikezelwa ngefomathi ye-Apache Parquet, kwaye ayizizo zonke iindibaniselwano zePython kunye nepakethi yeparquet esebenza ngaphandle kweempazamo. Ubunzima besibini yayikukukhuphela imifanekiso kwilifu lemeyile; okwangoku akukho ndlela ilula yokukhuphela isixa esikhulu sedatha ngexesha elinye. Ngenxa yoko, ezi ngxaki zilibazise abathathi-nxaxheba kangangeentsuku ezimbalwa.

IDAO. Inqanaba lokuqala

Umsebenzi wawukukwahlula i-muon / non-muon particles ngokweempawu zabo. Into ephambili yalo msebenzi yayibukho bekholomu yobunzima kwidatha yoqeqesho, apho abaququzeleli ngokwabo batolika njengokuzithemba kwimpendulo yalo mgca. Ingxaki yayikukuba imiqolo embalwa yayinobunzima obungalunganga.

Emva kokucinga imizuzu embalwa malunga nomgca kunye nelayini (ingcebiso itsalele ingqalelo kweli nqaku lekholamu yobunzima) kunye nokwakha le grafu, sigqibe kwelokuba sijonge iindlela ezi-3:

1) Jikela ekujoliswe kuko kwimigca enobunzima obungalunganga (kunye neentsimbi ngokufanelekileyo)
2) shenxisa iintsimbi ukuya kwelona xabiso lisezantsi ukuze ziqale ku-0
3) musa ukusebenzisa iintsimbi zobunzima

Inketho yesithathu yajika yaba yimbi kakhulu, kodwa ezimbini zokuqala zaphucula umphumo, eyona nto yayiyi-Nombolo ye-1, eyasizisa ngokukhawuleza kwindawo yesibini yangoku kumsebenzi wokuqala kunye neyokuqala kweyesibini.

Inyathelo lethu elilandelayo yayikukuphonononga idatha yamaxabiso angekhoyo. Abaququzeleli basinike idatha esele idibene, apho kwakukho amaxabiso ambalwa alahlekileyo, kwaye athatyathelwa indawo ngu -9999.

Sifumene amaxabiso angekhoyo kwiMatchedHit_{X,Y,Z}[N] kunye neMatchedHit_D{X,Y,Z}[N] ikholamu, kwaye kuphela xa N=2 okanye 3. Njengoko siqonda, amanye amasuntswana awazange dlula zonke ii-detectors ezi-4, kwaye wayeka nokuba kwi-plate yesi-3 okanye yesi-4. Idata ibiqulathe imiqolo yeLextra_{X,Y}[N], echaza into enye neMatchedHit_{X,Y,Z}[N], kodwa isebenzisa uhlobo oluthile lokudityaniswa. Olu qikelelo luncinci lucebise ukuba i-Lextra_{X,Y}[N] inokubekwa endaweni yamaxabiso angekhoyo kwi-MatchedHit_{X,Y,Z}[N] (ye X kunye no-Y banxibelelanisa kuphela). I-MatchedHit_Z[N] yayizaliswe kakuhle yi-median. Ezi nkohliso zisivumele ukuba sifikelele kwindawo yoku-1 ephakathi kuyo yomibini le misebenzi.

Ukuqwalasela ukuba abazange banike nantoni na ukuphumelela kwinqanaba lokuqala, sinokuyeka apho, kodwa saqhubeka, sakha imifanekiso emihle kwaye saza nezinto ezintsha.

Umzekelo, sifumanise ukuba xa sicwangcisa iindawo zokuhlangana zesuntswana ngalinye kwiipleyiti ezine zomtshini, sinokubona ukuba amanqaku akwiplate nganye adityaniswe abe ngamaxande ama-5 anomlinganiselo we-4 ukuya ku-5 kwaye agxininiswe kwindawo. inqaku (0,0), kwaye kwi Akukho manqaku kuxande lokuqala.

Inombolo yepleyiti / imilinganiselo yoxande	1	2	3	4	5
Icwecwe loku-1	500x625	1000x1250	2000x2500	4000x5000	8000x10000
Icwecwe loku-2	520x650	1040x1300	2080x2600	4160x5200	8320x10400
Icwecwe loku-3	560x700	1120x1400	2240x2800	4480x5600	8960x11200
Icwecwe loku-4	600x750	1200x1500	2400x3000	4800x6000	9600x12000

Emva kokumisela le milinganiso, songeze iimpawu ezi-4 ezintsha zecategorical kwisuntswana ngalinye - inani loxande apho linqumla ipleyiti nganye.

Siphinde saqaphela ukuba amaqhekeza abonakala esasaza emacaleni ukusuka kwiziko kwaye kwavela ingcamango yokuvavanya ngandlela-thile "umgangatho" wale ngqungquthela. Ngokufanelekileyo, mhlawumbi kuya kwenzeka ukuba kufike uhlobo oluthile lweparabola "efanelekileyo" ngokuxhomekeke kwindawo yokusuka kwaye uqikelele ukutenxa kuyo, kodwa sizikhawulele kumgca "ofanelekileyo" othe ngqo. Emva kokwakha imigca echanekileyo enjalo kwindawo nganye yokungena, siye sakwazi ukubala ukutenxa okusemgangathweni kwetrajectory yeqhekeza ngalinye ukusuka kulo mgca othe ngqo. Ekubeni umndilili wokutenxa kwithagethi = 1 yayiyi-152, kwaye kwithagethi = 0 yayingu-390, siye savavanya olu phawu njengelungileyo. Kwaye eneneni, eli nqaku lenze ngokukhawuleza ukuba libe phezulu kwezona ziluncedo.

Sivuyiswe kwaye songeza ukuphambuka kwazo zonke iindawo ze-4 ze-intersection kwi-particle nganye ukusuka kumgca ochanekileyo ochanekileyo njengeempawu ezi-4 ezongezelelweyo (kwaye nazo zasebenza kakuhle).

Ukudibanisa kumanqaku enzululwazi ngesihloko sokhuphiswano, esinikwe ngabaququzeleli, kwabangela ingcamango yokuba sikude neyokuqala ukusombulula le ngxaki kwaye, mhlawumbi, kukho uhlobo oluthile lwesofthiwe ekhethekileyo. Emva kokuba sifumene indawo yokugcina kwi-github apho iindlela ze-IsMuonSimple, i-IsMuon, i-IsMuonLoose ziphunyeziwe, sizidlulisele kwindawo yethu ngokuguqulwa okuncinci. Iindlela ngokwazo zazilula kakhulu: umzekelo, ukuba amandla angaphantsi komda othile, ngoko akusiyo i-muon, ngaphandle koko i-muon. Ezo mpawu zilula ngokucacileyo azikwazanga ukunika ukonyuka kwimeko yokusebenzisa i-gradient boosting, ngoko ke songeze omnye "umgama" obalulekileyo kumda. Ezi mpawu nazo ziphuculwe kancinci. Mhlawumbi, ngokuhlalutya iindlela ezikhoyo ngokucokisekileyo, kwakunokwenzeka ukufumana iindlela ezinamandla kwaye zongeze kwiimpawu.

Ekupheleni kokhuphiswano, siye salungisa kancinane isisombululo "sokukhawuleza" kwingxaki yesibini; ekugqibeleni, yahluke kwisiseko kula manqaku alandelayo:

1. Kwimiqolo enobunzima obuphakathi koko bekujoliswe kuko kuguqulwe
2. Kuzaliswe amaxabiso angekhoyo kwi-MatchedHit_{X,Y,Z}[N]
3. Ubunzulu buncitshisiwe ukuya ku-7
4. Ukunciphisa izinga lokufunda ukuya ku-0.1 (yayiyi-0.19)

Ngenxa yoko, sizame iimpawu ezingakumbi (ezingaphumelelanga kakhulu), iiparamitha ezikhethiweyo kunye ne-catboost eqeqeshiweyo, i-lightgbm kunye ne-xgboost, sizame ukuxutywa kweengqikelelo ezahlukeneyo kwaye ngaphambi kokuvula okwabucala siphumelele ngokuzithemba kumsebenzi wesibini, kwaye okokuqala sasiphakathi iinkokeli.

Emva kokuvula indawo yangasese sasikwindawo ye-10 kumsebenzi we-1 kunye ne-3 kwisibini. Zonke iinkokeli zaxubana, kwaye isantya ngasese sasiphezulu kune-libboard. Kubonakala ngathi idatha ibingalungiswanga kakuhle (okanye umzekelo bekungekho miqolo enezisindo ezingalunganga ngasese) kwaye oku bekucaphukisa kancinci.

I-SNA Hackathon 2019 - Imibhalo. Inqanaba lokuqala

Umsebenzi yayikukubeka izithuba zomsebenzisi kwinethiwekhi yoluntu ye-Odnoklassniki ngokusekwe kumbhalo oqulethweyo; Ukongeza kwisicatshulwa, bekukho ezinye iimpawu ezimbalwa zeposti (ulwimi, umnini, umhla kunye nexesha lokudala, umhla kunye nexesha lokujonga. ).

Njengeendlela zeklasiki zokusebenza ngesicatshulwa, ndingagxininisa iinketho ezimbini:

1. Ukwenza imephu yegama ngalinye kwisithuba se-n-dimensional vector kangangokuba amagama afanayo aneevektha ezifanayo (funda ngakumbi kwi inqaku lethu), ngoko mhlawumbi ufumane igama eliqhelekileyo lokubhaliweyo okanye ukusebenzisa iindlela ezithatha ingqalelo indawo ehambelanayo yamagama (CNN, LSTM/GRU).
2. Ukusebenzisa iimodeli ezinokusebenza ngokukhawuleza kunye nezivakalisi ezipheleleyo. Ngokomzekelo, uBert. Ngokwethiyori, le ndlela kufuneka isebenze ngcono.

Ekubeni oku kwakungamava am okuqala ngeetekisi, kuya kuba yinto ephosakeleyo ukufundisa umntu, ngoko ke ndiya kuzifundisa. Nanga amacebiso endizakuzinika wona ekuqaleni kokhuphiswano:

1. Ngaphambi kokuba ubaleke ukufundisa into, jonga idatha! Ukongeza kumbhalo ngokwawo, idatha yayineekholamu ezininzi kwaye kwakunokwenzeka ukucofa kakhulu kubo kunokuba ndenze. Eyona nto ilula kukwenza i-encoding ekujoliswe kuyo kwezinye iikholamu.
2. Musa ukufunda kuyo yonke idatha! Kwakukho idatha eninzi (malunga ne-17 yezigidi zemiqolo) kwaye kwakungekho mfuneko kwaphela ukuzisebenzisa zonke ukuvavanya iingcamango. Uqeqesho kunye nokulungiswa kwangaphambili bekucotha kakhulu, kwaye ngokuqinisekileyo ngendibenalo ixesha lokuvavanya iingqikelelo ezinomdla ngakumbi.
3. <Ingcebiso ephikisanayo> Akukho mfuneko yokukhangela imodeli yombulali. Ndachitha ixesha elide ndicinga u-Elmo noBert, ngethemba lokuba baya kundithatha ngokukhawuleza bandise kwindawo ephakamileyo, kwaye ngenxa yoko ndasebenzisa i-FastText yokufakela kwangaphambili ukuqeqeshwa kolwimi lwesiRashiya. Andikwazanga ukufikelela kwisantya esingcono kunye no-Elmo, kwaye bendingenalo ixesha lokuyiqikelela noBert.
4. <Ingcebiso ephikisanayo> Akukho mfuneko yokujonga into enye yombulali. Ukujonga idatha, ndiqaphele ukuba malunga neepesenti ezi-1 zeetekisi aziqulathanga isicatshulwa! Kodwa kwakukho amakhonkco kwezinye izixhobo, kwaye ndabhala i-parser elula eyavula indawo kwaye yakhupha isihloko kunye nenkcazo. Kwakubonakala ngathi licebo elilungileyo, kodwa ke ndathwalwa ndaza ndagqiba kwelokuba ndicazulule onke amakhonkco azo zonke iitekisi kwaye ndaphinda ndaphulukana nexesha elininzi. Konke oku khange kubonelele ngophuculo olubalulekileyo kwisiphumo sokugqibela (nangona ndiye ndacinga ukuvela, umzekelo).
5. Iimpawu zeClassic ziyasebenza. Thina kuGoogle, umzekelo, "iimpawu zombhalo kaggle", sifunde kwaye songeze yonke into. I-TF-IDF ibonelele ngophuculo, ngokunjalo neempawu zobalo ezifana nobude bombhalo, amagama, kunye nobungakanani beziphumlisi.
6. Ukuba kukho iikholamu zeDateTime, kufanelekile ukuba zicazululwe zibe ziimpawu ezininzi ezahlukeneyo (iiyure, iintsuku zeveki, njl.njl.). Yeyiphi imiba ekufuneka igxininiswe kufuneka ihlalutywe kusetyenziswa iigrafu/iimetriki ezithile. Apha, ngokukhawuleza, ndenze yonke into ngokuchanekileyo kwaye ndigxininise iimpawu eziyimfuneko, kodwa uhlalutyo oluqhelekileyo aluyi kuba buhlungu (umzekelo, njengoko senzile ekugqibeleni).

Ngenxa yolu khuphiswano, ndaqeqesha enye imodeli yeekeras ngokuguqulelwa kwamagama, kunye nenye esekelwe kwiLSTM kunye neGRU. Bobabini basebenzise i-FastText eqeqeshelwe kwangaphambili i-FastText yolwimi lwesiRashiya (ndazama ezinye izinto ezifakelweyo, kodwa ezi zizona zisebenze kakuhle). Emva kokulinganisa uqikelelo, ndithathe indawo yesi-7 yokugqibela kubathathi-nxaxheba abangama-76.

Emva kwenqanaba lokuqala lapapashwa inqaku nguNikolai Anokhin, owathatha indawo yesibini (wathatha inxaxheba ngaphandle kokhuphiswano), kunye nesisombululo sakhe ukuya kwinqanaba elithile waphinda wam, kodwa waya phambili ngenxa query-key-value ingqalelo indlela.

Inqanaba lesibini Kulungile & IDAO

Izigaba zesibini zokhuphiswano zenzeke phantse ngokulandelelana, ngoko ke ndagqiba ekubeni ndibajonge kunye.

Okokuqala, mna kunye neqela elitsha elifunyenweyo saphela kwiofisi ekhangayo yenkampani ye-Mail.ru, apho umsebenzi wethu wawukudibanisa iimodeli zeengoma ezintathu ukusuka kwinqanaba lokuqala - isicatshulwa, imifanekiso kunye nentsebenziswano. Kwabelwa oku ngaphezu kweentsuku ezi-2, eziye zajika zaba zincinci kakhulu. Enyanisweni, siye sakwazi ukuphinda iziphumo zethu ukusuka kwinqanaba lokuqala ngaphandle kokufumana nayiphi na inzuzo ekudibaniseni. Ekugqibeleni, sithathe indawo ye-5, kodwa asikwazanga ukusebenzisa imodeli yombhalo. Emva kokujonga izisombululo zabanye abathathi-nxaxheba, kubonakala ngathi kwakufanelekile ukuzama ukudibanisa iitekisi kunye nokongeza kwimodeli yokubambisana. Impembelelo yecala yale nqanaba yayiyimibono emitsha, ukudibana kunye nokunxibelelana nabathathi-nxaxheba abapholileyo kunye nabaququzeleli, kunye nokungabikho okunzima kokulala, okunokuthi kuchaphazele umphumo wesigaba sokugqibela se-IDAO.

Umsebenzi kwi-IDAO 2019 Inqanaba lokugqibela lalikuqikelela ixesha lokulinda lomyalelo wabaqhubi beeteksi zeYandex kwisikhululo sezindiza. Kwinqanaba lesi-2, imisebenzi emi-3 = izikhululo zeenqwelo-moya ezi-3 zichongiwe. Kwisikhululo seenqwelomoya ngasinye, idatha yomzuzu ngomzuzu malunga nenani leeodolo zeeteksi zeenyanga ezintandathu inikwe. Kwaye njengedatha yovavanyo, idatha yenyanga elandelayo kunye nomzuzu-ngomzuzu wedatha yee-odolo zeeveki ezi-2 ezidlulileyo zanikwa. Kwakukho ixesha elincinane (iintsuku ezi-1,5), umsebenzi wawucacile, umntu omnye kuphela ovela kwiqela weza kukhuphiswano - kwaye ngenxa yoko, yayiyindawo ebuhlungu ukuya ekupheleni. Iingcamango ezinomdla zazibandakanya iinzame zokusebenzisa idatha yangaphandle: imozulu, i-traffic jams kunye nezibalo ze-Yandex zomyalelo wetekisi. Nangona abaququzeleli bengakhange batsho ukuba zeziphi ezi zikhululo zeenqwelo moya, abaninzi abathathi-nxaxheba bacinga ukuba yiSheremetyevo, iDomodedovo kunye neVnukovo. Nangona le ngcinga yaphikiswa emva kokhuphiswano, iimpawu, umzekelo, ukusuka kwidatha yemozulu yaseMoscow ziphucule iziphumo zombini ekuqinisekiseni nakwibhodi yabaphambili.

isiphelo

1. Ukhuphiswano lweML lupholile kwaye lunomdla! Apha uya kufumana ukusetyenziswa kwezakhono kuhlalutyo lwedatha, kunye neemodeli ezinobuqili kunye nobuchule, kunye nengqiqo elula yamkelekile.
2. I-ML sele ingumthombo omkhulu wolwazi obonakala ngathi ukhula ngokukhawuleza. Ndizibekele usukelo lokuqhelana nemimandla eyahlukeneyo (imiqondiso, imifanekiso, iitafile, isicatshulwa) kwaye sele ndiqonda ukuba kuninzi kangakanani ukufunda. Umzekelo, emva kolu khuphiswano ndaye ndagqiba ekubeni ndifunde: ukudibanisa ialgorithms, ubuchule obuphambili bokusebenza ngamathala eencwadi okukhulisa igradient (ngakumbi, ukusebenza neCatBoost kwiGPU), uthungelwano lwe-capsule, indlela yokuhoya i-query-key-value.
3. Hayi ngokukaggle yedwa! Maninzi amanye amakhuphiswano apho kulula ukufumana ubuncinci isikipa, kwaye kukho amathuba amaninzi amanye amabhaso.
4. Nxibelelana! Sele kukho uluntu olukhulu kummandla wokufunda koomatshini kunye nohlalutyo lwedatha, kukho amaqela anomxholo kwi-telegram, i-slack, kunye nabantu abanzulu abavela kwi-Mail.ru, i-Yandex kunye nezinye iinkampani ziphendula imibuzo kunye noncedo lwabaqalayo kunye nabaqhubeka nendlela yabo kule ndawo. ulwazi.
5. Ndicebisa wonke umntu ophefumlelwe ngongoma yangaphambili ukuba andwendwele idathafest - inkomfa enkulu yamahhala eMoscow, eya kuba ngoMeyi 10-11.

umthombo: www.habr.com