Nkhaniyi ili ndi magawo awiri:
- Kufotokozera mwachidule za zomangamanga zina zamakina kuti zizindikire zinthu muzithunzi ndi magawo azithunzi okhala ndi maulalo omveka bwino azinthu kwa ine. Ndinayesa kusankha mafotokozedwe a kanema ndipo makamaka mu Russian.
- Gawo lachiwiri ndikuyesa kumvetsetsa momwe chitukuko cha ma neural network architectures. Ndipo matekinoloje ozikidwa pa iwo.
Chithunzi 1 - Kumvetsetsa ma neural network architectures sikophweka
Zonse zidayamba ndikupanga ma demo awiri amtundu wa chinthu ndikuzindikira pa foni ya Android:
- , data ikasinthidwa pa seva ndikutumizidwa ku foni. Gulu la zithunzi za mitundu itatu ya zimbalangondo: zofiirira, zakuda ndi teddy.
- pamene deta kukonzedwa pa foni palokha. Kuzindikira zinthu (kuzindikira chinthu) zamitundu itatu: hazelnuts, nkhuyu ndi masiku.
Pali kusiyana pakati pa ntchito zamagulu azithunzi, kuzindikira kwa chinthu mu chithunzi ndi . Chifukwa chake, pakufunika kudziwa kuti ndi ma neural network architecture ati omwe amazindikira zinthu pazithunzi ndi zomwe zingagawidwe. Ndapeza zitsanzo zotsatirazi za zomangamanga zomwe zili ndi maulalo omveka bwino azinthu kwa ine:
- Zomangamanga zingapo zochokera ku R-CNN (Rzigawo ndi Convolution Neural Nmawonekedwe): R-CNN, Fast R-CNN, , . Kuti muzindikire chinthu chomwe chili pachithunzi, mabokosi omangirira amaperekedwa pogwiritsa ntchito makina a Region Proposal Network (RPN). Poyambirira, njira yofufuzira pang'onopang'ono idagwiritsidwa ntchito m'malo mwa RPN. Kenako madera ochepera osankhidwa amadyetsedwa kuti alowetse ma neural network wamba kuti agawidwe. Zomangamanga za R-CNN zili ndi malupu omveka bwino a "za" m'madera ochepa, okwana 2000 akuyenda kudzera pa intaneti ya AlexNet. Malupu omveka bwino a "for" amachepetsa liwiro la kukonza zithunzi. Kuchuluka kwa malupu omveka bwino omwe akudutsa mu neural network yamkati kumachepa ndi mtundu uliwonse watsopano wa zomangamanga, ndipo zosintha zina zambiri zimapangidwira kuti ziwonjezeke liwiro ndikusintha ntchito yozindikira chinthu ndikugawa chinthu mu Mask R-CNN.
- (You Only Look Once) ndiye neural network yoyamba yomwe idazindikira zinthu munthawi yeniyeni pazida zam'manja. Mawonekedwe apadera: kusiyanitsa zinthu pamayendedwe amodzi (ingoyang'anani kamodzi). Ndiko kuti, mu zomangamanga za YOLO mulibe malupu omveka bwino a "for", chifukwa chake maukonde amagwira ntchito mofulumira. Mwachitsanzo, fanizo ili: mu NumPy, pochita ntchito ndi matrices, palibenso malupu omveka bwino a "for", omwe mu NumPy akugwiritsidwa ntchito pamagulu otsika a zomangamanga kudzera m'chinenero cha pulogalamu ya C. YOLO imagwiritsa ntchito gridi ya mazenera otchulidwatu. Kuletsa chinthu chomwecho kuti chisatanthauzidwe kangapo, zenera la overlap coefficient (IoU) imagwiritsidwa ntchito. Imphambano oVesi Union). Zomangamangazi zimagwira ntchito mosiyanasiyana ndipo zimakhala zapamwamba : Chitsanzo chikhoza kuphunzitsidwa pazithunzi koma chimagwirabe bwino pa zojambula zojambula pamanja.
- (Skubuula Sotentha MultiBox Detector) - "ma hacks" opambana kwambiri a zomangamanga za YOLO amagwiritsidwa ntchito (mwachitsanzo, kuponderezedwa kosalekeza) ndipo zatsopano zimawonjezeredwa kuti neural network igwire ntchito mofulumira komanso molondola. Chinthu chosiyana: kusiyanitsa zinthu mumayendedwe amodzi pogwiritsa ntchito gridi yopatsidwa ya windows (bokosi lokhazikika) pa piramidi ya chithunzi. Piramidi ya chithunziyo imayikidwa mu ma convolution tensor kudzera motsatizanatsatizana ndi ntchito zophatikizira (ndi ntchito yophatikiza kwambiri, kukula kwa malo kumachepa). Mwanjira imeneyi, zinthu zazikulu ndi zazing'ono zimatsimikiziridwa mumayendedwe amodzi.
- MobileSSD (mafoniNetV2+ SSD) ndi kuphatikiza kwa ma neural network architectures awiri. Network yoyamba imagwira ntchito mwachangu ndikuwonjezera kulondola kwa kuzindikira. MobileNetV2 imagwiritsidwa ntchito m'malo mwa VGG-16, yomwe idagwiritsidwa ntchito poyambirira . Network yachiwiri ya SSD imatsimikizira malo omwe zinthu zili pachithunzichi.
- - neural network yaying'ono koma yolondola. Payokha, sichithetsa vuto la kuzindikira zinthu. Komabe, itha kugwiritsidwa ntchito pophatikiza mamangidwe osiyanasiyana. Ndipo amagwiritsidwa ntchito pazida zam'manja. Chodziwika bwino ndichakuti deta imayamba kuphatikizidwa kukhala zosefera zinayi za 1 Γ 1 ndikukulitsidwa kukhala zosefera zinayi za 1 Γ 1 ndi zinayi 3 Γ 3 zosefera. Kuchulukitsa kotere kwa kufalikira kwa data kumatchedwa "Fire Module".
- (Semantic Image Segmentation with Deep Convolutional Nets) - magawo a zinthu pachithunzichi. Chosiyana ndi kamangidwe kameneka ndi dilated convolution, yomwe imateteza kusasunthika kwa malo. Izi zimatsatiridwa ndi gawo lokonzekera zotsatira pogwiritsa ntchito chitsanzo cha graphical probabilistic (conditional random field), chomwe chimakulolani kuchotsa phokoso laling'ono m'magulu ndikuwongolera khalidwe lachithunzi chogawidwa. Kumbuyo kwa dzina lochititsa chidwi la "graphical probabilistic model" kumabisala fyuluta wamba ya Gaussian, yomwe ili pafupi ndi mfundo zisanu.
- Anayesa kupeza chipangizo (Kuwombera Kumodzi Sinthayambitsani Neural Network for Object Theection), koma sindinamve zambiri.
- Ndidawonanso momwe ukadaulo wa "tcheru" umagwirira ntchito: , , . Chodziwika bwino pakupanga "chidwi" ndikusankha kwapadera kwa zigawo zomwe zimachulukitsidwa pachithunzicho (RoI, Rzigawo of Iinterest) pogwiritsa ntchito neural network yotchedwa Attention Unit. Madera omwe ali ndi chidwi chowonjezereka amafanana ndi mabokosi omangirira, koma mosiyana ndi iwo, samakhazikika pachithunzichi ndipo akhoza kukhala ndi malire osamveka. Kenako, kuchokera kumadera omwe akuchulukirachulukira, zizindikilo (ziwonetsero) zimapatulidwa, zomwe "zimadyetsedwa" ku neural network zokhazikika ndi zomanga. . Ma neural network obwereza amatha kusanthula ubale wazinthu motsatizana. Ma neural network omwe amapezeka pafupipafupi adagwiritsidwa ntchito kumasulira mawu m'zilankhulo zina, ndipo tsopano kumasulira ΠΈ .
Pamene tikufufuza zomanga izi Ndinazindikira kuti sindikumvetsa kalikonse. Ndipo sikuti ma neural network anga ali ndi vuto ndi makina osamala. Kulengedwa kwa zomangamanga zonsezi kuli ngati mtundu wina wa hackathon yaikulu, kumene olemba amapikisana mu hacks. Kuthyolako ndi njira yachangu yothetsera vuto zovuta mapulogalamu. Ndiko kuti, palibe mgwirizano wowoneka ndi womveka womveka pakati pa zomangamanga zonsezi. Zonse zomwe zimawagwirizanitsa ndi gulu la ma hacks opambana kwambiri omwe amabwereka wina ndi mzake, kuphatikizapo wamba kwa onse. (kubweza zolakwika, kubweza kumbuyo). Ayi ! Sizikudziwika bwino zomwe zingasinthidwe komanso momwe mungakwaniritsire zomwe zilipo kale.
Chifukwa cha kusowa kwa kulumikizana pakati pa ma hacks, ndizovuta kwambiri kukumbukira ndikugwiritsa ntchito pochita. Ichi ndi chidziwitso chogawikana. Zabwino kwambiri, nthawi zingapo zosangalatsa komanso zosayembekezereka zimakumbukiridwa, koma zambiri zomwe zimamveka komanso zosamvetsetseka zimasowa kukumbukira mkati mwa masiku angapo. Zidzakhala bwino ngati mu sabata mumakumbukira osachepera dzina la zomangamanga. Koma maola angapo ngakhale masiku ogwirira ntchito anathera poΕ΅erenga nkhani ndi kuonera mavidiyo obwereza!
Chithunzi 2 -
Olemba ambiri a nkhani za sayansi, m'malingaliro anga, amachita zonse zotheka kuti atsimikizire kuti ngakhale chidziwitso chogawikachi sichimveka ndi owerenga. Koma mawu otenga nawo gawo m'mizere khumi yokhala ndi ma formula omwe amatengedwa "kuchokera mumlengalenga" ndi mutu wankhani ina (vuto ).
Pazifukwa izi, pakufunika kukonza zidziwitso pogwiritsa ntchito ma neural network, motero, kukulitsa kumvetsetsa komanso kuloweza pamtima. Chifukwa chake, mutu waukulu wakuwunika matekinoloje amunthu payekha komanso kamangidwe ka ma neural network ochita kupanga unali ntchito iyi: fufuzani kumene izo zonse zikupita, osati chipangizo cha neural network iliyonse payokha.
Kodi zonsezi zikupita kuti? Zotsatira zazikulu:
- Chiwerengero cha makina oyambira ophunzirira zaka ziwiri zapitazi . Chifukwa chotheka: "ma neural network salinso chatsopano."
- Aliyense akhoza kupanga neural network yogwira ntchito kuti athetse vuto losavuta. Kuti muchite izi, tengani chitsanzo chokonzekera kuchokera ku "model zoo" ndikuphunzitsa gawo lomaliza la neural network () pa data yopangidwa kale kuchokera kapena kuchokera mu mfulu .
- Opanga akuluakulu a neural network adayamba kupanga "malo osungira nyama" (chitsanzo zoo). Pogwiritsa ntchito mutha kupanga pulogalamu yamalonda mwachangu: kwa TensorFlow, kwa PyTorch, kwa Caffe2, kwa Chainer ndi .
- Neural network ikugwira ntchito pompopompo (nthawi yeniyeni) pazida zam'manja. Kuyambira 10 mpaka 50 mafelemu pamphindikati.
- Kugwiritsa ntchito ma neural network mu mafoni (TF Lite), mu asakatuli (TF.js) ndi mu (IoT, Ichakumadzulo of Tmatumba). Makamaka m'mafoni omwe amathandizira kale ma neural network pamlingo wa hardware (neural accelerators).
- βChida chilichonse, zovala, ngakhale chakudya chidzakhala nacho IP-v6 adilesi ndikulumikizana wina ndi mnzake" - .
- Chiwerengero cha zofalitsa pa kuphunzira pamakina chayamba kukula (kuwirikiza kawiri zaka ziwiri zilizonse) kuyambira 2015. Zachidziwikire, timafunikira ma neural network kuti tifufuze zolemba.
- Tekinoloje zotsatirazi zikutchuka:
- PyTorch - kutchuka kukukulirakulira ndipo zikuwoneka kuti zikupitilira TensorFlow.
- Zosankha zokha za hyperparameter Chizindikiro cha AutoML - kutchuka kukukula bwino.
- Kutsika pang'onopang'ono pakulondola ndikuwonjezeka kwa liwiro la kuwerengera: , ma aligorivimu , kuwerengera kosawerengeka (pafupifupi), quantization (pamene zolemera za neural network zimasinthidwa kukhala integers ndi quantized), neural accelerators.
- Kutanthauzira ΠΈ .
- chilengedwe , tsopano mu nthawi yeniyeni.
- Chinthu chachikulu chokhudza DL ndi chakuti pali deta yambiri, koma kusonkhanitsa ndi kulemba sikophweka. Chifukwa chake, markup automation ikukula () kwa neural network pogwiritsa ntchito neural network.
- Ndi ma neural network, Computer Science idakhala mwadzidzidzi sayansi yoyesera ndipo adawuka .
- Ndalama za IT komanso kutchuka kwa ma neural network kudawonekera nthawi imodzi pomwe kompyuta idakhala mtengo wamsika. Chuma chikusintha kuchoka ku chuma cha golide ndi ndalama kupita ku golide-ndalama-computing. Onani nkhani yanga ndi chifukwa cha maonekedwe a IT ndalama.
Pang'onopang'ono china chatsopano chikuwonekera (Kuphunzira Kwamakina & Kuphunzira Mwakuya), zomwe zakhazikika pakuyimira pulogalamuyi ngati magulu ophunzitsidwa bwino a neural network.
Chithunzi 3 - ML / DL monga njira yatsopano yopangira mapulogalamu
Komabe, sizinawonekere "Neural network theory", momwe mungaganizire ndikugwira ntchito mwadongosolo. Zomwe tsopano zimatchedwa "nthanthi" kwenikweni ndizoyesera, ma aligorivimu anzeru.
Maulalo kuzinthu zanga ndi zina:
- Nkhani ya Sayansi ya Data. Makamaka kukonza zithunzi. Aliyense amene akufuna kulandira ayenera kutumiza imelo (foobar167<gaf-gaf>gmail<dot>com). Ndimatumiza maulalo azolemba ndi makanema pomwe zinthu zikuwunjikana.
- Zambiri zomwe ndadutsa ndi zomwe ndikufuna kuzidutsa.
- , komwe muyenera kuyamba kuphunzira ma neural network. Plus kabuku .
- , kumene aliyense adzapeza chinachake chosangalatsa kwa iyemwini.
- Tinawaona kukhala othandiza kwambiri. makanema owunikira zolemba zasayansi by Data Science. Pezani, lembetsani kwa iwo ndikupereka maulalo kwa anzako ndi inenso. Zitsanzo:
- aka ndi malangizo a tsatane-tsatane ndi gwero lotseguka.
Zikomo chifukwa cha chidwi chanu!
Source: www.habr.com