ProHoster > Π‘Π»ΠΎΠ³ > xov xwm hauv internet > Neural networks. Qhov no mus qhov twg?

Neural networks. Qhov no mus qhov twg?

Kab lus muaj ob ntu:

  1. Cov lus piav qhia luv luv ntawm qee lub network architectures rau kev tshawb pom cov khoom hauv cov duab thiab cov duab segmentation nrog cov kev nkag siab zoo tshaj plaws rau cov peev txheej rau kuv. Kuv sim xaiv cov lus piav qhia video thiab nyiam dua hauv Lavxias.
  2. Qhov thib ob yog ib qho kev sim nkag siab txog kev txhim kho ntawm neural network architectures. Thiab technologies raws li lawv.

Neural networks. Qhov no mus qhov twg?

Daim duab 1 - Kev nkag siab txog neural network architectures tsis yooj yim

Nws tag nrho pib los ntawm kev ua ob daim ntawv thov demo rau kev faib khoom thiab tshawb pom ntawm lub xov tooj Android:

  • Back-end demo, thaum cov ntaub ntawv ua tiav ntawm lub server thiab xa mus rau lub xov tooj. Kev faib duab ntawm peb hom bears: xim av, dub thiab teddy.
  • Pem hauv ntej-kawg demothaum cov ntaub ntawv ua tiav ntawm lub xov tooj nws tus kheej. Kev txheeb xyuas cov khoom (cov khoom tshawb pom) ntawm peb hom: hazelnuts, figs thiab hnub tim.

Nws muaj qhov sib txawv ntawm cov dej num ntawm kev faib cov duab, nrhiav pom cov khoom hauv ib qho duab thiab duab segmentation. Yog li ntawd, yuav tsum tau nrhiav kom paub seb qhov twg neural network architectures ntes cov khoom hauv cov duab thiab qhov twg tuaj yeem ua ntu. Kuv pom cov piv txwv hauv qab no ntawm architectures nrog cov kev nkag siab tshaj plaws rau cov peev txheej rau kuv:

  • A series ntawm architectures raws li R-CNN (Rcheeb tsam nrog Ckev vam meej Neural Networks nta): R-CNN, Ceev R-CNN, Faster R-CNN, Mask R-CNN. Txhawm rau txhawm rau txheeb xyuas cov khoom hauv ib daim duab, cov thawv ntawv khij tau faib siv Regional Proposal Network (RPN) mechanism. Thaum pib, qhov kev tshawb nrhiav qeeb qeeb tau siv los ntawm RPN. Tom qab ntawd xaiv cov cheeb tsam txwv tau pub rau cov tswv yim ntawm cov pa neural network rau kev faib tawm. R-CNN architecture tau qhia meej "rau" lub voj voog hla thaj tsam txwv, tag nrho txog 2000 khiav los ntawm AlexNet sab hauv network. Nthuav "rau" loops qeeb cov duab ua kom ceev. Tus naj npawb ntawm cov loops qhia tau meej khiav los ntawm lub hauv nruab nrab neural network txo qis nrog txhua tus tshiab version ntawm lub architecture, thiab kaum ob ntawm lwm yam kev hloov tau ua kom ceev thiab hloov cov hauj lwm ntawm cov khoom nrhiav pom nrog cov khoom segmentation hauv Mask R-CNN.
  • YOLO (You Only Look Once) yog thawj lub network neural uas lees paub cov khoom hauv lub sijhawm tiag tiag ntawm cov khoom siv txawb. Distinctive feature: qhov txawv ntawm cov khoom hauv ib qho kev khiav (tsuas yog saib ib zaug). Ntawd yog, hauv YOLO architecture tsis muaj qhov qhia meej "rau" loops, uas yog vim li cas lub network ua haujlwm sai. Piv txwv li, qhov kev sib piv no: hauv NumPy, thaum ua haujlwm nrog matrices, kuj tsis muaj "rau" loops, uas nyob rau hauv NumPy yog siv nyob rau theem qis ntawm cov architecture los ntawm C programming lus. Txhawm rau tiv thaiv tib yam khoom los ntawm kev txhais ntau zaus, lub qhov rais sib tshooj coefficient (IoU) yog siv. Ikev sib tshuam oVer Unco). Qhov no architecture ua hauj lwm nyob rau hauv ib tug ntau yam thiab muaj siab kev ruaj khov: Tus qauv tuaj yeem cob qhia cov duab tab sis tseem ua tau zoo ntawm cov duab kos duab.
  • SSD (Spuab tais SKub MultiBox Detector) - qhov ua tau zoo tshaj plaws "hacks" ntawm YOLO architecture yog siv (piv txwv li, tsis muaj kev txwv ntau tshaj) thiab cov tshiab ntxiv los ua kom cov neural network ua haujlwm sai dua thiab raug. Distinctive feature: qhov txawv ntawm cov khoom nyob rau hauv ib tug khiav siv ib tug muab daim phiaj ntawm qhov rais (default kem) ntawm daim duab pyramid. Cov duab pyramid yog encoded nyob rau hauv convolution tensors los ntawm successive convolution thiab pooling ua hauj lwm (nrog rau lub max-pooling lag luam, lub spatial dimension txo). Ua li no, ob qho tib si loj thiab me me yog txiav txim siab hauv ib lub network khiav.
  • MobileSSD (mobileNetV2+ SSD) yog kev sib txuas ntawm ob lub neural network architectures. Thawj lub network MobileNetV2 ua haujlwm sai thiab ua kom paub tseeb qhov tseeb. MobileNetV2 yog siv los hloov VGG-16, uas yog thawj zaug siv hauv thawj tsab ntawv. Qhov thib ob SSD network txiav txim qhov chaw ntawm cov khoom hauv daim duab.
  • SqueezeNet - ib qho me me tab sis muaj tseeb neural network. Los ntawm nws tus kheej, nws tsis daws qhov teeb meem ntawm kev tshawb nrhiav khoom. Txawm li cas los xij, nws tuaj yeem siv rau hauv kev sib xyaw ntawm cov qauv sib txawv. Thiab siv nyob rau hauv mobile pab kiag li lawm. Qhov txawv feature yog tias cov ntaub ntawv yog thawj zaug compressed rau hauv plaub 1 Γ— 1 convolutional lim thiab ces nthuav mus rau plaub 1 Γ— 1 thiab plaub 3 Γ— 3 convolutional lim. Ib qho iteration ntawm cov ntaub ntawv compression-expansion yog hu ua "Fire Module".
  • DeepLab (Semantic Image Segmentation with Deep Convolutional Nets) - segmentation ntawm cov khoom hauv daim duab. Ib tug txawv feature ntawm lub architecture yog dilated convolution, uas khaws cia spatial daws teeb meem. Qhov no yog ua raws li cov theem tom qab ua tiav ntawm cov txiaj ntsig uas siv cov qauv graphical probabilistic (qhov xwm txheej random teb), uas tso cai rau koj tshem tawm cov suab nrov me me hauv segmentation thiab txhim kho qhov zoo ntawm cov duab segmented. Tom qab lub npe hu ua "graphical probabilistic model" hides cov pa lim Gaussian, uas yog kwv yees los ntawm tsib lub ntsiab lus.
  • Sim ua kom paub cov cuab yeej RefineDet (Single-Shot Caisment Neural Network for Object nwsection), tab sis kuv tsis nkag siab ntau.
  • Kuv kuj tau saib yuav ua li cas "kev saib xyuas" technology ua haujlwm: video 1, video 2, video 3. Ib qho tshwj xeeb ntawm "kev saib xyuas" architecture yog qhov kev xaiv tsis siv neeg ntawm thaj chaw ntawm kev saib xyuas hauv daim duab (RoI, Rlegions of Interest) siv lub neural network hu ua Attention Unit. Cov cheeb tsam ntawm kev mloog ntau ntxiv zoo ib yam li cov thawv khi, tab sis tsis zoo li lawv, lawv tsis raug kho nyob rau hauv daim duab thiab tej zaum yuav muaj qhov muag plooj. Tom qab ntawd, los ntawm thaj tsam ntawm kev mloog ntau ntxiv, cov paib (cov yam ntxwv) raug cais tawm, uas yog "nyuaj" rau cov kev sib txuas ntawm cov neural nrog cov architectures. LSDM, GRU los yog Vanilla RNN. Recurrent neural networks muaj peev xwm txheeb xyuas kev sib raug zoo ntawm cov yam ntxwv hauv ib ntu. Cov kev sib txuas ntawm cov neural tau pib siv los txhais cov ntawv ua lwm hom lus, thiab tam sim no rau kev txhais lus duab rau ntawv ΠΈ ntawv rau duab.

Thaum peb tshawb nrhiav cov architectures no Kuv pom tau tias kuv tsis nkag siab dab tsi. Thiab nws tsis yog tias kuv lub network neural muaj teeb meem nrog cov txheej txheem saib xyuas. Lub creation ntawm tag nrho cov architectures yog zoo li ib yam ntawm cov loj loj hackathon, qhov twg sau ntawv sib tw nyob rau hauv hacks. Hack yog ib qho kev daws sai rau qhov teeb meem software nyuaj. Ntawd yog, tsis muaj qhov pom thiab nkag siab qhov kev sib txuas ntawm txhua qhov kev tsim qauv no. Txhua yam uas koom ua ke lawv yog cov txheej txheem ntawm kev vam meej tshaj plaws hacks uas lawv qiv los ntawm ib leeg, ntxiv rau ib qho rau txhua tus kaw-loop convolution ua haujlwm (yuam kev backpropagation, backpropagation). Tsis muaj kev xav! Nws tsis paub meej tias yuav hloov dab tsi thiab yuav ua li cas txhawm rau txhim kho kev ua tiav uas twb muaj lawm.

Raws li qhov tshwm sim ntawm qhov tsis muaj kev sib txuas ntawm kev sib txuas ntawm hacks, lawv nyuaj heev kom nco ntsoov thiab siv rau hauv kev xyaum. Qhov no yog fragmented kev paub. Qhov zoo tshaj plaws, ob peb lub sijhawm ntxim nyiam thiab tsis xav txog yuav nco qab, tab sis feem ntau ntawm qhov nkag siab thiab nkag siab tsis tau ploj ntawm kev nco hauv ob peb hnub. Nws yuav zoo yog tias nyob rau hauv ib lub lis piam koj nco ntsoov tsawg kawg yog lub npe ntawm lub architecture. Tab sis ob peb teev thiab cov hnub ua haujlwm tau siv sijhawm nyeem cov ntawv thiab saib cov yeeb yaj kiab tshuaj xyuas!

Neural networks. Qhov no mus qhov twg?

Daim duab 2 – Zoo ntawm Neural Networks

Feem ntau cov kws sau ntawv ntawm cov ntawv tshawb fawb, hauv kuv tus kheej lub tswv yim, ua txhua yam ua tau los xyuas kom meej tias txawm tias qhov kev paub tsis meej no tsis nkag siab los ntawm tus nyeem ntawv. Tab sis cov kab lus koom tes hauv kaum kab kab lus nrog cov qauv uas tau muab "tawm ntawm huab cua" yog ib lub ntsiab lus rau ib tsab xov xwm (teeb meem tshaj tawm lossis piam sij).

Vim li no, nws yog ib qho tsim nyog yuav tsum tau ua kom cov ntaub ntawv siv cov neural networks thiab, yog li, nce kev nkag siab zoo thiab kev nco. Yog li ntawd, lub ntsiab lus tseem ceeb ntawm kev tsom xam ntawm tus kheej technologies thiab architectures ntawm artificial neural networks yog cov haujlwm hauv qab no: nrhiav seb nws yuav mus qhov twg, thiab tsis yog lub cuab yeej ntawm ib qho tshwj xeeb neural network cais.

Txhua yam no mus qhov twg? Cov txiaj ntsig tseem ceeb:

  • Tus naj npawb ntawm kev pib kawm tshuab hauv ob xyoos dhau los poob siab. Yog vim li cas: "neural networks tsis yog ib yam tshiab lawm."
  • Leej twg tuaj yeem tsim lub network neural ua haujlwm los daws qhov teeb meem yooj yim. Txhawm rau ua qhov no, coj tus qauv npaj los ntawm "tus qauv zoo" thiab cob qhia cov txheej txheem kawg ntawm neural network (hloov kev kawm) ntawm cov ntaub ntawv npaj tau los ntawm Google Dataset Nrhiav lossis ntawm 25 txhiab Kaggle datasets hauv dawb huab Jupyter Notebook.
  • Cov tuam txhab loj ntawm neural networks pib tsim "model zoo" (model zoo). Siv lawv koj tuaj yeem tsim daim ntawv thov kev lag luam sai: TF Hub rau TensorFlow, MMDetection rau PyTorch, Ntes rau Caffe2, chainer-modelzoo rau Chainer thiab lwm yam.
  • Neural networks ua haujlwm hauv sijhawm tiag (real-time) ntawm cov khoom siv mobile. Los ntawm 10 mus rau 50 ntas ib ob.
  • Kev siv neural networks hauv xov tooj (TF Lite), hauv browsers (TF.js) thiab hauv cov khoom siv hauv tsev (IoT, Iinternet of Things). Tshwj xeeb tshaj yog nyob rau hauv cov xov tooj uas twb txhawb neural tes hauj lwm ntawm lub hardware theem (neural accelerators).
  • β€œTxhua yam khoom siv, khoom siv khaub ncaws, thiab tej zaum txawm tias zaub mov yuav muaj IP-v6 chaw nyob thiab sib txuas lus nrog ib leeg" - Sebastian Thrun.
  • Tus naj npawb ntawm cov ntawv tshaj tawm ntawm kev kawm tshuab tau pib loj tuaj tshaj Moore txoj cai (ob zaug txhua ob xyoos) txij li xyoo 2015. Obviously, peb xav tau neural networks rau kev txheeb xyuas cov khoom.
  • Cov thev naus laus zis hauv qab no tau txais koob meej:
    • PyTorch - Kev muaj koob meej loj hlob sai thiab zoo li tau dhau TensorFlow.
    • Tsis siv neeg xaiv ntawm hyperparameters AutoML - muaj koob meej yog loj hlob zoo.
    • Maj mam txo qhov raug thiab nce hauv kev suav ceev: fuzzy logic, cov algorithms txhawb nqa, inexact (kwv yees) xam, quantization (thaum lub luj ntawm lub neural network hloov dua siab tshiab rau integers thiab quantized), neural accelerators.
    • Hloov Mus duab rau ntawv ΠΈ ntawv rau duab.
    • creation XNUMXD khoom los ntawm video, tam sim no hauv lub sijhawm tiag tiag.
    • Qhov tseem ceeb ntawm DL yog tias muaj ntau cov ntaub ntawv, tab sis kev sau thiab sau npe nws tsis yooj yim. Yog li ntawd, markup automation tab tom txhim kho (automated annotation) rau neural networks siv neural networks.
  • Nrog neural tes hauj lwm, Computer Science dheev los ua kev tshawb fawb thiab sawv kev rov tsim dua tshiab.
  • IT nyiaj thiab qhov nrov ntawm neural networks tshwm sim ib txhij thaum xam los ua tus nqi lag luam. Kev lag luam hloov pauv los ntawm kev lag luam kub thiab txiaj mus rau kub-txiaj-xws li. Saib kuv tsab xov xwm ntawm econophysics thiab yog vim li cas rau qhov tsos ntawm IT nyiaj.

Maj mam ib qho tshiab tshwm ML/DL programming methodology (Machine Learning & Deep Learning), uas yog raws li sawv cev rau qhov kev pab cuam raws li ib txheej ntawm kev cob qhia neural network qauv.

Neural networks. Qhov no mus qhov twg?

Daim duab 3 – ML/DL raws li ib tug tshiab programming methodology

Txawm li cas los xij, nws yeej tsis tshwm sim "neural network txoj kev xav", nyob rau hauv uas koj tuaj yeem xav thiab ua haujlwm tau zoo. Tam sim no hu ua "kev xav" yog qhov kev sim, heuristic algorithms.

Txuas rau kuv thiab lwm yam kev pab:

Ua tsaug rau koj txoj kev paub!

Tau qhov twg los: www.hab.com

Ntxiv ib saib