Amanethiwekhi e-Neural. Kuyaphi konke lokhu?

I-athikili inezingxenye ezimbili:

1. Incazelo emfushane yezinye izakhiwo zenethiwekhi zokutholwa kwento ezithombeni nokuhlukaniswa kwesithombe ngezixhumanisi eziqondakala kakhulu zezinsiza kimi. Ngizamile ukukhetha izincazelo zevidiyo futhi okungcono ngesiRashiya.
2. Ingxenye yesibili umzamo wokuqonda isiqondiso sokuthuthukiswa kwe-neural network architectures. Nobuchwepheshe obusekelwe kubo.

Umfanekiso 1 - Ukuqonda izakhiwo zenethiwekhi ye-neural akulula

Konke kwaqala ngokwenza izinhlelo zokusebenza ezimbili zedemo zokuhlelwa kwento nokutholwa ocingweni lwe-Android:

• Idemo yangemuva, uma idatha icutshungulwa kuseva futhi idluliselwa ocingweni. Ukuhlukaniswa kwesithombe kwezinhlobo ezintathu zamabhere: ansundu, amnyama kanye ne-teddy.
• Idemo yangaphambililapho idatha icutshungulwa ocingweni ngokwalo. Ukutholwa kwezinto (ukutholwa kwezinto) zezinhlobo ezintathu: ama-hazelnuts, amakhiwane kanye nezinsuku.

Kunomehluko phakathi kwemisebenzi yokuhlukaniswa kwesithombe, ukutholwa kwento esithombeni kanye ukuhlukaniswa kwesithombe. Ngakho-ke, bekunesidingo sokuthola ukuthi yimaphi ama-neural network architecture athola izinto ezithombeni nokuthi yiziphi ezingahlukanisa. Ngithole izibonelo ezilandelayo zezakhiwo ezinezixhumanisi eziqondakala kakhulu zezinsiza kimi:

• Uchungechunge lwezakhiwo ezisekelwe ku-R-CNN (Rizifunda nge Convolution Neural Nizici zokusebenza): R-CNN, Fast R-CNN, I-R-CNN esheshayo, Imaski R-CNN. Ukuze kutholwe into esesithombeni, amabhokisi ahlanganisayo anikezwa kusetshenziswa indlela ye-Regional Proposal Network (RPN). Ekuqaleni, kwasetshenziswa indlela yokusesha ehamba kancane ekhethiwe esikhundleni se-RPN. Bese izifunda ezinomkhawulo ezikhethiwe zinikezwa okokufaka kwenethiwekhi ye-neural evamile ukuze ihlukaniswe. Isakhiwo se-R-CNN sinezihibe ezicacile "ze" ezindaweni ezilinganiselwe, ezifinyelela ku-2000 zigijima kunethiwekhi yangaphakathi ye-AlexNet. Izihibe ezicacile "ze" zehlisa isivinini sokucubungula isithombe. Inani lamaluphu acacile agijima kunethiwekhi ye-neural yangaphakathi liyehla ngenguqulo entsha ngayinye yezakhiwo, futhi inqwaba yezinye izinguquko ziyenziwa ukuze kukhuliswe isivinini futhi kuthathelwe indawo umsebenzi wokutholwa kwento ngokuhlukaniswa kwento kuMask R-CNN.
• YOLO (You Only Look Once) iyinethiwekhi yokuqala ye-neural eyabona izinto ngesikhathi sangempela kumadivayisi eselula. Isici esihlukile: izinto ezihlukanisayo ngokugijima okukodwa (bheka nje kanye). Okusho ukuthi, ekwakhiweni kwe-YOLO azikho izihibe ezicacile "ze", yingakho inethiwekhi isebenza ngokushesha. Isibonelo, lesi sifaniso: ku-NumPy, uma wenza imisebenzi ngo-matrices, awekho futhi amaluphu “wa” acacile, okuthi ku-NumPy asetshenziswe emazingeni aphansi ezakhiwo ngolimi lokuhlela C. I-YOLO isebenzisa igridi yamawindi achazwe ngaphambilini. Ukuze uvimbele into efanayo ukuthi ingachazwa izikhathi eziningi, kusetshenziswa i-coefficient (IoU) yokugqagqana kwewindi. Iimpambanamgwaqo oNhlobo Union). Lesi sakhiwo sisebenza ebangeni elibanzi futhi sinokuphezulu ukuqina: Imodeli ingaqeqeshwa ezithombeni kodwa isasebenza kahle emidwebeni edwetshwe ngesandla.
• I-SSD (Subuhlalu Si-MultiBox eshisayo Di-etector) - "ama-hacks" aphumelele kakhulu we-architecture ye-YOLO asetshenziswa (isibonelo, ukucindezelwa okungeyona okukhulu) futhi okusha kwengezwa ukuze kwenziwe inethiwekhi ye-neural isebenze ngokushesha nangaphezulu. Isici esihlukile: izinto ezihlukanisayo ngokugijima okukodwa kusetshenziswa igridi enikeziwe yamawindi (ibhokisi elimisiwe) kuphiramidi yesithombe. Iphiramidi yesithombe ifakwe ikhodi ku-convolution tensor ngokusebenzisa i-convolution elandelanayo kanye nemisebenzi yokuhlanganisa (ngomsebenzi wokuhlanganisa okuphezulu, ubukhulu bendawo buyancipha). Ngale ndlela, kokubili izinto ezinkulu nezincane zinqunywa ekugijimeni kwenethiwekhi eyodwa.
• I-MobileSSD (HambayoI-NetV2+ I-SSD) iyinhlanganisela yezakhiwo ezimbili zenethiwekhi ye-neural. Inethiwekhi yokuqala I-MobileNetV2 isebenza ngokushesha futhi ikhulisa ukunemba kokuqashelwa. I-MobileNetV2 isetshenziswa esikhundleni se-VGG-16, eyayisetshenziswa ekuqaleni isihloko sokuqala. Inethiwekhi yesibili ye-SSD inquma indawo yezinto ezisesithombeni.
• I-SqueezeNet - inethiwekhi ye-neural encane kakhulu kodwa enembile. Ngokwayo, ayixazululi inkinga yokutholwa kwento. Noma kunjalo, ingasetshenziswa ekuhlanganiseni kwezakhiwo ezahlukene. Futhi isetshenziswa kumadivayisi eselula. Isici esihlukile ukuthi idatha iqala icindezelwe ibe izihlungi ezine ze-1×1 bese inwetshwa ibe izihlungi ezine ze-1×1 nezine ze-3×3 zokuguqula. Okunye okunjalo kokucindezelwa kwedatha kubizwa ngokuthi “Imojula Yomlilo”.
• I-DeepLab (Isegmentation yesithombe se-Semantic enamanethi e-Deep Convolutional) - ukuhlukaniswa kwezinto ezisesithombeni. Isici esihlukile sesakhiwo i-convolution enwetshiwe, egcina ukulungiswa kwendawo. Lokhu kulandelwa yisigaba sangemva kokucubungula semiphumela kusetshenziswa imodeli yegraphical probabilistic (inkambu engahleliwe enemibandela), ekuvumela ukuthi ukhiphe umsindo omncane ekuhlukaniseni futhi uthuthukise ikhwalithi yesithombe esihlukanisiwe. Ngemuva kwegama elesabekayo elithi "graphical probabilistic model" kufihla isihlungi esivamile se-Gaussian, esilinganiswa ngamaphoyinti amahlanu.
• Uzame ukuthola idivayisi RefineDet (Isibhamu Sinye HlelaiNeural Network for Object Theection), kodwa angizange ngiqonde okuningi.
• Ngiphinde ngabheka ukuthi ubuchwepheshe "bokunaka" busebenza kanjani: ividiyo1, ividiyo2, ividiyo3. Isici esihlukile sesakhiwo "sokunaka" ukukhetha okuzenzakalelayo kwezifunda zokunakwa okwengeziwe esithombeni (i-RoI, Ramabutho of Iinterest) usebenzisa inethiwekhi ye-neural ebizwa ngokuthi i-Attention Unit. Izifunda zokunakwa okwengeziwe zifana namabhokisi abophayo, kodwa ngokungafani nazo, azinqunyelwe esithombeni futhi zingase zibe nemingcele efiphele. Khona-ke, kusukela ezifundeni zokunakwa okwengeziwe, izimpawu (izici) ziyahlukaniswa, "zondliwa" kumanethiwekhi we-neural aphindaphindiwe anezakhiwo. I-LSDM, i-GRU noma i-Vanilla RNN. Amanethiwekhi we-neural ajwayelekile akwazi ukuhlaziya ubudlelwano bezici ngokulandelana. Amanethiwekhi avamile e-neural ekuqaleni asetshenziselwa ukuhumushela umbhalo kwezinye izilimi, futhi manje ukuze ahunyushwe izithombe kumbhalo и umbhalo uye esithombeni.

Njengoba sihlola lezi zakhiwo Ngabona ukuthi angiqondi lutho. Futhi akukhona ukuthi inethiwekhi yami ye-neural inezinkinga ngendlela yokunaka. Ukudalwa kwazo zonke lezi zakhiwo kufana nohlobo oluthile lwe-hackathon enkulu, lapho ababhali beqhudelana khona ngama-hack. I-Hack yisisombululo esisheshayo senkinga yesofthiwe enzima. Okusho ukuthi, akukho ukuxhumana okunengqondo okubonakalayo nokuqondakalayo phakathi kwazo zonke lezi zakhiwo. Konke okubahlanganisayo isethi yama-hack aphumelele kakhulu abawaboleka komunye nomunye, kanye nokujwayelekile kwabo bonke umsebenzi wokuguqula iluphu evaliwe (i-backpropagation yephutha, i-backpropagation). Cha izinhlelo zokucabanga! Akukacaci ukuthi yini okufanele ishintshwe nendlela yokuthuthukisa izimpumelelo ezikhona.

Njengomphumela wokuntuleka kokuxhumana okunengqondo phakathi kwama-hacks, kunzima kakhulu ukuwakhumbula nokusebenzisa ekusebenzeni. Lolu ulwazi oluyizicucu. Okungcono kakhulu, izikhathi ezimbalwa ezithakazelisayo nezingalindelekile zikhunjulwa, kodwa okuningi kwalokho okuqondwayo nokungaqondakali kuyanyamalala enkumbulweni phakathi nezinsuku ezimbalwa. Kuyoba kuhle uma ngesonto ukhumbula okungenani igama le-architecture. Kodwa amahora amaningana ngisho nezinsuku zokusebenza zachithwa kufundwa izihloko nokubuka amavidiyo okubuyekeza!

Umfanekiso 2 - I-Zoo ye-Neural Networks

Iningi lababhali bezihloko zesayensi, ngokombono wami siqu, benza konke okusemandleni ukuqinisekisa ukuthi ngisho nalolu lwazi oluyizicucu aluqondwa ngumfundi. Kodwa imishwana ebambe iqhaza emishweni yemigqa eyishumi enamafomula athathwe “emoyeni omncane” iyisihloko sendatshana ehlukile (inkinga ukushicilela noma ukushabalala).

Ngalesi sizathu, kunesidingo sokuhlela ulwazi kusetshenziswa amanethiwekhi e-neural futhi, ngaleyo ndlela, kwandise ikhwalithi yokuqonda nokubamba ngekhanda. Ngakho-ke, isihloko esiyinhloko sokuhlaziywa kobuchwepheshe obubodwa kanye nezakhiwo zamanethiwekhi e-neural okwenziwa kwaba umsebenzi olandelayo: thola ukuthi konke kuyaphi, futhi hhayi idivayisi yanoma iyiphi inethiwekhi ye-neural ngokuhlukana.

Kuyaphi konke lokhu? Imiphumela esemqoka:

• Inani lokuqalisa ukufunda komshini eminyakeni emibili edlule wawa kakhulu. Isizathu esingaba khona: "amanethiwekhi e-neural awaseyona into entsha."
• Noma ubani angakha inethiwekhi ye-neural esebenzayo ukuze axazulule inkinga elula. Ukuze wenze lokhu, thatha imodeli eseyenziwe ngomumo "ezoo eyimodeli" futhi uqeqeshe ungqimba lokugcina lwenethiwekhi ye-neural (ukudlulisa ukufunda) kudatha esenziwe ngomumo evela Usesho lwe-Google Dataset noma kusuka Izinkulungwane ezingama-25 zedathasethi ye-Kaggle mahhala ifu Jupyter Notebook.
• Abakhiqizi abakhulu bamanethiwekhi e-neural baqala ukudala "ama-zoo amamodeli" (isibonelo sezu). Ukuzisebenzisa ungakha ngokushesha uhlelo lokusebenza lwezentengiso: Ihabhu le-TF okwe-TensorFlow, I-MMDetection ye-PyTorch, I-Detectron yeCaffe2, Chainer-modelzoo ngoba Chainer futhi другие.
• Amanethiwekhi e-Neural asebenza kuwo isikhathi sangempela (isikhathi sangempela) kumadivayisi eselula. Kusukela ku-10 kuya ku-50 ozimele ngomzuzwana.
• Ukusetshenziswa kwamanethiwekhi e-neural kumafoni (TF Lite), kuziphequluli (TF.js) kanye naku izinto zasendlini (IoT, Iisibongo of Tizintambo). Ikakhulukazi kumafoni asevele esekela amanethiwekhi e-neural ezingeni lehadiwe (ama-neural accelerators).
• Zonke izinto, izingubo, mhlawumbe ngisho nokudla kuzoba nakho Ikheli le-IP-v6 nokuxhumana ”- Sebastian Thrun.
• Inombolo yokushicilelwa kokufundwa komshini isiqalile ukukhula yeqa umthetho kaMoore (iphindwe kabili njalo eminyakeni emibili) kusukela ngo-2015. Ngokusobala, sidinga amanethiwekhi e-neural ukuze sihlaziye izindatshana.
• Ubuchwepheshe obulandelayo buzuza ukuduma:
  • I-PyTorch - ukuduma kukhula ngokushesha futhi kubonakala kuyidlula i-TensorFlow.
  • Ukukhetha okuzenzakalelayo kwama-hyperparameter I-AutoML - ukuthandwa kukhula kahle.
  • Ukwehla kancane kancane kokunemba kanye nokwenyuka kwesivinini sokubala: logic engaqondakali, ama-algorithms ukuqinisa, izibalo ezingaqondile (okungenani), ukulinganisa (uma izisindo zenethiwekhi ye-neural ziguqulwa zibe izinombolo eziphelele futhi zilinganiswe), ama-neural accelerators.
  • Ukuhumusha izithombe kumbhalo и umbhalo uye esithombeni.
  • indalo Izinto ze-3D ezivela kuvidiyo, manje ngesikhathi sangempela.
  • Into eyinhloko nge-DL ukuthi kunedatha eminingi, kodwa ukuqoqa nokulebula akulula. Ngakho-ke, i-markup automation iyathuthuka (isichasiselo esizenzakalelayo) kumanethiwekhi e-neural asebenzisa amanethiwekhi e-neural.
• Ngamanethiwekhi e-neural, iComputer Science yaba ngokuzumayo isayensi yokuhlola wasukuma inkinga yokukhiqiza kabusha.
• Imali ye-IT kanye nokuduma kwamanethiwekhi e-neural kwavela ngesikhathi esisodwa lapho ikhompuyutha iba inani lemakethe. Umnotho uyashintsha usuka kumnotho wegolide nowemali ukuya gold-currency-computing. Bheka isihloko sami ku i-econophysics kanye nesizathu sokuvela kwemali ye-IT.

Kancane kancane kuvela entsha Indlela yokuhlela ye-ML/DL (Ukufunda Ngomshini nokufunda Okujulile), okusekelwe ekumeleleni uhlelo njengesethi yamamodeli enethiwekhi ye-neural aqeqeshiwe.

Umfanekiso 3 – ML/DL njengendlela entsha yokuhlela

Nokho, ayizange ibonakale "ithiyori yenethiwekhi ye-neural", lapho ungacabanga futhi usebenze ngokuhlelekile. Lokho manje okubizwa ngokuthi “ithiyori” empeleni kuwukuhlola, ama-algorithms we-heuristic.

Izixhumanisi zami nezinye izinsiza:

• Iphephandaba Isayensi Yedatha. Ikakhulukazi ukucutshungulwa kwesithombe. Noma ubani ofuna ukuyithola kufanele athumele i-imeyili (foobar167gmailcom). Ngithumela izixhumanisi ezihlokweni nakumavidiyo njengoba kuqoqwana.
• Okujwayelekile uhlu lwezifundo nezindatshanaengidlulile futhi engingathanda ukudlula.
• Izifundo namavidiyo wabaqalayo, lapho kufanele uqale khona ukufunda amanethiwekhi we-neural. Plus incwajana "Isingeniso Sokufunda Ngomshini kanye Namanethiwekhi Okwenziwa Kwezinzwa".
• Amathuluzi Awusizo, lapho wonke umuntu ezothola okuthile okuthakazelisayo kubo.
• Siwathole ewusizo kakhulu. iziteshi zevidiyo zokuhlaziya izindatshana zesayensi by Isayensi Yedatha. Thola, ubhalisele kubo futhi udlulisele izixhumanisi kozakwenu kanye nami futhi. Izibonelo:
  • Amaphepha Amabili Wemizuzu
  • Henry AI Labs
  • Yannic Kilcher
  • IkhodiEmporium
  • Ibhulogi Chengwei Zhang aka Tony607 ngemiyalo yesinyathelo ngesinyathelo kanye nomthombo ovulekile.

Спасибо за внимание!

Source: www.habr.com