Labarin ya ƙunshi sassa biyu:
- Takaitaccen bayanin wasu gine-ginen cibiyar sadarwa don gano abu a cikin hotuna da rarrabuwar hoto tare da mafi fahimtar hanyoyin haɗin kai zuwa albarkatu a gare ni. Na yi ƙoƙarin zaɓar bayanin bidiyo kuma zai fi dacewa a cikin Rashanci.
- Kashi na biyu ƙoƙari ne na fahimtar alkiblar ci gaban gine-ginen hanyoyin sadarwa na jijiyoyi. Da kuma fasahohin da suka dogara da su.
Hoto 1 - Fahimtar gine-ginen cibiyar sadarwar jijiyoyi ba sauki ba
An fara duka ta hanyar yin aikace-aikacen demo guda biyu don tantancewa da gano abubuwa akan wayar Android:
- , lokacin da aka sarrafa bayanai akan uwar garken kuma ana aika su zuwa wayar. Rarraba hoto na nau'ikan bears guda uku: launin ruwan kasa, baki da teddy.
- lokacin da ake sarrafa bayanai akan wayar kanta. Gano abubuwa (gano abu) iri uku: hazelnuts, figs da dabino.
Akwai bambanci tsakanin ayyukan rarraba hoto, gano abu a cikin hoto da . Don haka, akwai buƙatar gano waɗanne gine-ginen cibiyar sadarwa na jijiyoyi ke gano abubuwa a cikin hotuna da waɗanne za su iya raba. Na sami misalan gine-gine masu zuwa tare da mafi fahimtar hanyoyin haɗin kai zuwa albarkatu a gare ni:
- Jerin gine-gine bisa R-CNN (Ryankuna da Cjuyin juya hali Neural NFasalolin etworks): R-CNN, Mai sauri R-CNN, , . Don gano abu a cikin hoto, ana keɓance akwatunan ɗaure ta amfani da tsarin Sadarwar Sadarwar Yanki (RPN). Da farko, an yi amfani da tsarin Neman Zaɓa a hankali maimakon RPN. Sannan ana ciyar da yankuna masu iyaka da aka zaɓa zuwa shigar da hanyar sadarwa ta al'ada don rarrabuwa. Gine-gine na R-CNN yana da madaidaicin madaukai "don" akan yankuna masu iyaka, jimlar har zuwa 2000 yana gudana ta hanyar sadarwar ciki ta AlexNet. Madaidaitan madaukai "don" suna rage saurin sarrafa hoto. Adadin madaukai bayyanannun da ke gudana ta hanyar hanyar sadarwa ta jijiyoyi na ciki yana raguwa tare da kowane sabon sigar gine-gine, kuma ana yin ɗimbin sauran canje-canje don ƙara saurin gudu da maye gurbin aikin gano abu tare da ɓarna abu a cikin Mask R-CNN.
- (You Only Look Once) ita ce hanyar sadarwa ta farko wacce ta gane abubuwa a ainihin lokacin akan na'urorin hannu. Siffa ta musamman: rarrabe abubuwa a cikin gudu ɗaya (duba sau ɗaya kawai). Wato, a cikin tsarin gine-gine na YOLO babu takamaiman "don" madaukai, wanda shine dalilin da ya sa cibiyar sadarwa ke aiki da sauri. Misali, wannan kwatankwacin: a cikin NumPy, lokacin yin ayyuka tare da matrices, kuma babu takamaiman “don” madaukai, waɗanda a cikin NumPy ana aiwatar da su a ƙananan matakan gine-gine ta harshen shirye-shiryen C. YOLO yana amfani da grid na windows da aka riga aka ƙayyade. Don hana fayyace abu iri ɗaya sau da yawa, ana amfani da haɗin haɗin gwiwar tagar (IoU). Imahada oVer Ununi). Wannan gine-ginen yana aiki a kan kewayon da yawa kuma yana da girma : Ana iya horar da samfurin akan hotuna amma har yanzu yana da kyau akan zanen da aka zana.
- (Smakwancin gwaiwa Szafi MultiBox Detector) - ana amfani da "hacks" mafi nasara na tsarin gine-gine na YOLO (misali, rashin matsawa mafi girma) kuma an ƙara sababbin don sa cibiyar sadarwar jijiyoyi ta yi aiki da sauri da kuma daidai. Siffa ta musamman: rarrabe abubuwa a cikin gudu ɗaya ta amfani da grid ɗin windows (akwatin tsoho) akan hoton dala. Hoton dala yana kunshe ne a cikin tenors na juyin juya hali ta hanyar jujjuyawar juzu'i da ayyukan hadawa (tare da max-pooling, girman sararin samaniya yana raguwa). Ta wannan hanyar, duka manyan da ƙananan abubuwa ana ƙaddara su a cikin hanyar sadarwa guda ɗaya.
- MobileSSD (MobileNetV2+ SSD) hade ne na gine-ginen hanyoyin sadarwa na jijiyoyi guda biyu. Cibiyar sadarwa ta farko yana aiki da sauri kuma yana ƙara daidaiton ganewa. Ana amfani da MobileNetV2 maimakon VGG-16, wanda aka fara amfani dashi a ciki . Cibiyar sadarwa ta SSD ta biyu tana ƙayyade wurin abubuwan da ke cikin hoton.
- – ƙanƙanta amma ingantaccen hanyar sadarwa na jijiyoyi. Da kanta, ba ta magance matsalar gano abu ba. Duk da haka, ana iya amfani dashi a cikin haɗin gine-gine daban-daban. Kuma ana amfani dashi a cikin na'urorin hannu. Siffar ta musamman ita ce an fara matsa bayanan zuwa cikin matatun juyin juya hali na 1 × 1 sannan kuma a fadada su zuwa matatun juyin juya hali 1 × 1 da hudu 3 × 3. Ɗaya daga cikin irin wannan juzu'i na matsawa-faɗin bayanai ana kiransa "Fire Module".
- (Semantic Image Segmentation with Deep Convolutional Nets) - rarrabuwar abubuwa a cikin hoton. Musamman fasalin gine-ginen shine ruɗaɗɗen juzu'i, wanda ke adana ƙudurin sarari. Wannan yana biyo bayan matakin aiwatarwa na sakamakon ta amfani da ƙirar yuwuwar hoto (filin bazuwar yanayi), wanda ke ba ku damar cire ƙaramin ƙara a cikin sashin kuma inganta ingancin hoton da aka raba. Bayan ƙaƙƙarfan sunan “samfurin yuwuwar hoto” yana ɓoye matatar Gaussian ta al'ada, wacce ke kusan maki biyar.
- Kokarin gano na'urar (Single-Shot Ƙarfafament Neural Network for Abu Daection), amma ban fahimta da yawa ba.
- Na kuma duba yadda fasahar “hankali” ke aiki: , , . Wani fasali na musamman na gine-ginen "hankali" shine zaɓi na atomatik na yankuna na ƙarin kulawa a cikin hoton (RoI, Ral'adu of Isha'awa) ta amfani da hanyar sadarwa na jijiyar da ake kira Atention Unit. Yankunan da aka ƙara hankali sun yi kama da akwatuna masu iyaka, amma ba kamar su ba, ba a daidaita su a cikin hoton ba kuma suna iya zama masu iyakoki. Sa'an nan kuma, daga yankunan da aka kara da hankali, alamun (fasali) an ware su, waɗanda aka "ciyar da su" zuwa cibiyoyin sadarwa na yau da kullum tare da gine-gine. . Cibiyoyin jijiyoyi na yau da kullun suna iya yin nazarin dangantakar fasali a cikin jeri. An fara amfani da cibiyoyin sadarwa na yau da kullun don fassara rubutu zuwa wasu harsuna, yanzu kuma don fassarawa и .
Yayin da muke bincika waɗannan gine-ginen Na gane cewa ban gane komai ba. Kuma ba wai cibiyar sadarwa ta jijiyoyi na da matsaloli tare da tsarin kulawa ba. Ƙirƙirar duk waɗannan gine-ginen kamar wani nau'i ne na babban hackathon, inda marubuta ke gasa a hack. Hack shine mafita mai sauri ga matsalar software mai wahala. Wato babu wata alaka ta hankali da fahimta tsakanin dukkan wadannan gine-gine. Duk abin da ya haɗa su, saitin hacks ɗin da suka fi samun nasara waɗanda suke aro daga juna, da na gama-gari na kowa. (kuskure backpropagation, backpropagation). A'a ! Ba a bayyana abin da za a canza da yadda za a inganta nasarorin da ake samu ba.
Sakamakon rashin alaƙar ma'ana tsakanin hacks, suna da matukar wahalar tunawa da amfani a aikace. Wannan shi ne rarrabuwar kawuna. A mafi kyau, ana tunawa da wasu lokuta masu ban sha'awa da ba tsammani, amma yawancin abin da aka fahimta da rashin fahimta sun ɓace daga ƙwaƙwalwar ajiya a cikin 'yan kwanaki. Zai yi kyau idan a cikin mako guda ka tuna aƙalla sunan gine-gine. Amma an kashe sa'o'i da yawa har ma da kwanakin aiki don karanta labarai da kallon bidiyon bita!
Hoto na 2 -
Yawancin marubutan labarin kimiyya, a ra'ayi na, suna yin duk mai yiwuwa don tabbatar da cewa ko da wannan rarrabuwar ilimin ba mai karatu ya fahimce shi ba. Amma jumlolin tarayya a cikin jimlolin layi guda goma tare da dabarun da aka ɗauka "daga cikin iska" wani batu ne na wani labarin dabam (matsala). ).
Saboda wannan dalili, akwai buƙatar tsara bayanai ta amfani da hanyoyin sadarwa na jijiyoyi kuma, don haka, ƙara ingancin fahimta da haddacewa. Sabili da haka, babban jigon nazarin fasaha na mutum ɗaya da gine-ginen hanyoyin sadarwa na wucin gadi shine aiki mai zuwa: gano inda duk ya dosa, kuma ba na'urar kowane takamaiman hanyar sadarwa na jijiyoyi daban ba.
Ina duk wannan ke tafiya? Babban sakamako:
- Yawan fara koyon inji a cikin shekaru biyu da suka gabata . Dalili mai yuwuwa: "Cibiyoyin sadarwa na jijiyoyi ba sabon abu bane."
- Kowa na iya ƙirƙirar cibiyar sadarwar jijiyoyi mai aiki don magance matsala mai sauƙi. Don yin wannan, ɗauki samfurin da aka ƙera daga "zuwan samfurin" kuma horar da layin ƙarshe na cibiyar sadarwar jijiyoyi () akan bayanan da aka shirya daga ko daga a kyauta .
- Manyan masana'antun na cibiyoyin sadarwa na jijiyoyi sun fara ƙirƙirar "Zo model" (model zoo). Amfani da su zaku iya ƙirƙirar aikace-aikacen kasuwanci da sauri: don TensorFlow, don PyTorch, don Kafe2, ga Chainer da .
- Neural networks suna aiki a ciki ainihin lokaci (ainihin lokaci) akan na'urorin hannu. Daga firam 10 zuwa 50 a sakan daya.
- Amfani da hanyoyin sadarwa na jijiyoyi a cikin wayoyi (TF Lite), a cikin masu bincike (TF.js) da a ciki (IoT, Iyanar gizo of Thanta). Musamman a cikin wayoyi waɗanda suka riga sun goyi bayan hanyoyin sadarwa na jijiyoyi a matakin kayan aiki (masu hawan jini).
- "Kowane na'ura, kayan tufafi, da watakila ma abinci za su samu Adireshin IP-v6 kuma a sadarwa da juna" - .
- Adadin wallafe-wallafen kan koyon injin ya fara girma (biyu duk shekara biyu) tun 2015. Babu shakka, muna buƙatar hanyoyin sadarwar jijiya don nazarin labarai.
- Fasaha masu zuwa suna samun farin jini:
- PyTorch - shahararsa yana girma cikin sauri kuma da alama yana ƙetare TensorFlow.
- Zaɓin hyperparameters ta atomatik AutoML – shahararsa na girma lafiya.
- Rage daidaito a hankali da haɓaka saurin lissafi: , algorithms , Ƙididdigar ƙididdiga (kimanin) ƙididdiga, ƙididdigewa (lokacin da aka canza ma'auni na cibiyar sadarwar jijiyoyi zuwa ƙididdiga da ƙididdigewa), masu haɓaka jijiyoyi.
- Fassara и .
- halittar , yanzu a hakikanin lokaci.
- Babban abu game da DL shine cewa akwai bayanai da yawa, amma tattarawa da lakafta shi ba sauki. Don haka, ana samun ci gaba ta atomatik ta atomatik.) don hanyoyin sadarwar jijiyoyi ta amfani da hanyoyin sadarwa na jijiyoyi.
- Tare da hanyoyin sadarwa na jijiyoyi, Kimiyyar Kwamfuta ta zama kwatsam kimiyyar gwaji kuma ya tashi .
- Kuɗin IT da shaharar hanyoyin sadarwar jijiyoyi sun fito lokaci guda lokacin da kwamfuta ta zama darajar kasuwa. Tattalin arzikin yana canzawa daga tattalin arzikin zinari da kudin waje zuwa zinariya-kudin-kwamfuta. Dubi labarina akan da dalilin bayyanar kudin IT.
A hankali wata sabuwa ta bayyana (Machine Learning & Deep Learning), wanda ya dogara ne akan wakiltar shirin a matsayin saiti na ƙwararrun hanyoyin sadarwa na jijiyoyi.
Hoto 3 – ML/DL a matsayin sabuwar hanyar tsara shirye-shirye
Duk da haka, bai taba bayyana ba "ka'idar hanyar sadarwa ta jijiyoyi", a cikin abin da za ku iya tunani da aiki da tsari. Abin da ake kira yanzu "ka'idar" shine ainihin gwaji, algorithms heuristic.
Hanyoyin haɗi zuwa nawa da sauran albarkatu:
- Jaridar Kimiyyar Bayanai. Musamman sarrafa hoto. Duk wanda ke son karba sai ya aiko da sakon Imel (foobar167gmailcom). Ina aika hanyoyin haɗi zuwa labarai da bidiyo yayin da abubuwa ke taruwa.
- Janar wanda na wuce kuma wanda nake so in wuce.
- , inda ya kamata ka fara nazarin hanyoyin sadarwa na jijiyoyi. Ƙarin kasida .
- , inda kowa zai sami wani abu mai ban sha'awa ga kansa.
- Mun same su da amfani sosai. tashoshin bidiyo don nazarin labaran kimiyya by Data Science. Nemo, biyan kuɗi zuwa gare su kuma ku ba da hanyoyin haɗin kai ga abokan aikinku da ni ma. Misalai:
- aka tare da umarnin mataki-mataki da buɗe tushen.
Na gode da hankali!
source: www.habr.com