Neural networks. Qhov no mus qhov twg?

Kab lus muaj ob ntu:

1. Ib qho lus piav qhia luv luv txog qee cov duab nrhiav thiab faib cov qauv network, nrog rau cov kev sib txuas rau cov peev txheej uas kuv pom tias nkag siab zoo tshaj plaws. Kuv sim xaiv cov lus piav qhia video, nyiam dua hauv Lavxias.
2. Ntu thib ob sim nkag siab txog kev coj ua ntawm kev txhim kho ntawm cov qauv neural network thiab cov thev naus laus zis raws li lawv.

Daim Duab 1 - Kev nkag siab txog cov qauv ntawm lub network neural tsis yooj yim

Txhua yam pib nrog kuv ua ob lub demo apps rau kev faib khoom thiab kev tshawb nrhiav khoom ntawm kuv lub xov tooj. Android:

• Demo tom qab, thaum cov ntaub ntawv raug ua tiav ntawm lub server thiab xa mus rau lub xov tooj. Kev faib tawm duab ntawm peb hom dais: xim av, dub, thiab teddy.
• Kev ua qauv qhia pem hauv ntej, thaum cov ntaub ntawv raug ua tiav ntawm lub xov tooj nws tus kheej. Kev nrhiav pom khoom ntawm peb hom: txiv laum huab xeeb, txiv nkhaus taw, thiab hnub tim.

Muaj qhov sib txawv ntawm cov haujlwm faib duab, kev nrhiav khoom hauv daim duab, thiab kev faib cov duabYog li ntawd, nws tau los ua qhov tsim nyog los kawm seb cov qauv neural network twg ntes tau cov khoom hauv cov duab thiab qhov twg tuaj yeem faib lawv. Kuv pom cov piv txwv hauv qab no ntawm cov qauv nrog cov kev sib txuas nkag siab zoo tshaj plaws rau cov peev txheej:

• Ib qho ntawm cov qauv R-CNN-based (Rthaj tsam nrog Ckev hloov pauv Neural Ncov yam ntxwv ntawm etworks): R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNNYuav kom nrhiav tau cov khoom hauv ib daim duab, cov thawv ciam teb raug rho tawm siv lub Region Proposal Network (RPN) mechanism. Thaum xub thawj, lub tshuab Selective Search qeeb dua tau siv es tsis yog RPN. Cov thawv ciam teb raug rho tawm ces raug pub rau lub neural network ib txwm muaj rau kev faib tawm. Lub R-CNN architecture siv cov voj voog "rau" kom rov ua dua cov thawv ciam teb, nrog txog li 2000 khiav los ntawm lub network AlexNet sab hauv. Cov voj voog "rau" no ua rau qeeb qeeb rau kev ua cov duab. Tus lej ntawm cov voj voog thiab khiav los ntawm lub neural network sab hauv raug txo nrog txhua qhov version tshiab ntawm lub architecture, thiab ntau kaum lwm qhov kev hloov pauv tau ua kom nce qhov ceev thiab hloov txoj haujlwm ntawm kev nrhiav khoom nrog kev faib khoom hauv Mask R-CNN.
• YOLO (You Otsis muaj dab tsi Look Once) yog thawj lub neural network los paub txog cov khoom hauv lub sijhawm tiag tiag ntawm cov khoom siv mobile. Nws qhov tshwj xeeb yog lub peev xwm los paub qhov txawv ntawm cov khoom hauv ib qho kev khiav (tsuas yog ib qho saib). Ntawd yog, YOLO architecture tsis siv cov voj voog "rau" meej, uas tso cai rau lub network ua haujlwm sai. Piv txwv li, ntawm no yog qhov piv txwv: NumPy kuj tsis siv cov voj voog "rau" meej thaum ua cov haujlwm matrix; cov no tau siv rau hauv NumPy ntawm qib qis dua ntawm architecture los ntawm C programming lus. YOLO siv lub grid ntawm cov qhov rai uas tau teev tseg ua ntej. Txhawm rau tiv thaiv tib yam khoom los ntawm kev raug kuaj pom ntau zaus, qhov rais overlapping ratio (IoU) siv. Ikev sib tshuam oVer Union). Cov qauv no ua haujlwm ntau yam thiab muaj siab kev ruaj khov: Tus qauv tuaj yeem cob qhia rau ntawm cov duab, tab sis tseem ua haujlwm zoo rau ntawm cov duab kos tes.
• SSD (Slus Askiv SLub thawv ntau yam kub DLub YOLO neural network siv cov "hacks" zoo tshaj plaws ntawm YOLO architecture (piv txwv li, tsis yog qhov siab tshaj plaws suppression) thiab ntxiv cov tshiab kom ua rau nws sai dua thiab raug dua. Nws qhov tshwj xeeb yog lub peev xwm los sib txawv cov khoom hauv ib qho kev khiav siv lub grid ntawm cov qhov rai (lub thawv tsis siv neeg) ntawm lub pyramid duab. Lub pyramid duab yog encoded hauv convolutional tensors siv cov kev sib txuas convolution thiab pooling ua haujlwm (nrog max-pooling, qhov chaw dimensionality txo qis). Qhov no tso cai rau ob qho tib si loj thiab me me kom pom hauv ib qho kev khiav network.
• Txawb SSD (mobileNetV2 + SSD) yog kev sib xyaw ua ke ntawm ob lub neural network architectures. Lub network thawj zaug MobileNetV2 Nws ua haujlwm sai thiab ua rau kom paub tseeb dua. MobileNetV2 siv hloov VGG-16, uas tau siv thawj zaug hauv thawj tsab ntawvLub network SSD thib ob txiav txim siab qhov chaw ntawm cov khoom hauv daim duab.
• SqueezeNet - ib lub neural network me me tab sis raug. Nws tsis daws qhov teeb meem ntawm kev nrhiav khoom ntawm nws tus kheej. Txawm li cas los xij, nws tuaj yeem siv ua ke nrog ntau yam architectures thiab tsim nyog rau kev siv ntawm cov khoom siv mobile. Nws qhov tshwj xeeb yog tias cov ntaub ntawv raug compressed ua ntej rau hauv plaub 1 × 1 convolutional lim, tom qab ntawd nthuav mus rau hauv plaub 1 × 1 thiab plaub 3 × 3 convolutional lim. Ib qho kev rov ua dua ntawm cov ntaub ntawv compression thiab nthuav dav hu ua "Fire Module."
• DeepLab (Kev Faib Duab Semantic nrog Deep Convolutional Nets) - kev faib cov khoom hauv ib daim duab. Ib qho tshwj xeeb ntawm cov qauv yog dilated convolution, uas khaws cia qhov kev daws teeb meem spatial. Qhov no yog ua raws li theem tom qab ua tiav siv cov qauv graphical probability (ib qho conditional random field), uas tshem tawm cov suab nrov me me hauv kev faib thiab txhim kho qhov zoo ntawm daim duab segmented. Tom qab lub npe txaus ntshai "graphical probability model" muaj ib lub lim dej Gaussian yooj yim kwv yees los ntawm tsib lub ntsiab lus.
• Kuv sim nrhiav seb lub cuab yeej ntawd RefineDet (Ib Zaug Xwb Caiskev sib txuas lus Neural Network rau Object nwsection), tab sis nkag siab me ntsis.
• Kuv kuj tau saib seb cov thev naus laus zis "mloog" ua haujlwm li cas: video 1, video 2, video 3Ib qho tshwj xeeb ntawm "kev mloog zoo" architecture yog qhov kev xaiv tsis siv neeg ntawm cov cheeb tsam ntawm kev mloog zoo ntxiv hauv daim duab (RoI, Rlegions of Ikev txaus siab) siv lub neural network hu ua Attention Unit. Cov cheeb tsam mloog zoo ib yam li cov thawv bounding, tab sis tsis zoo li lawv, lawv tsis ruaj khov hauv daim duab thiab tuaj yeem muaj cov ciam teb tsis meej. Cov yam ntxwv tom qab ntawd raug rho tawm ntawm cov cheeb tsam mloog thiab pub rau hauv cov neural networks rov ua dua nrog cov architectures. LSDM, GRU, lossis Vanilla RNNCov tes hauj lwm neural uas rov tshwm sim tuaj yeem tshuaj xyuas kev sib raug zoo ntawm cov yam ntxwv hauv ib qho kev sib law liag. Cov tes hauj lwm neural uas rov tshwm sim tau siv rau kev txhais cov ntawv nyeem rau lwm hom lus, thiab tam sim no rau kev txhais lus cov duab rau cov ntawv nyeem и ntawv rau duab.

Thaum peb kawm txog cov architectures no, Kuv pom tias kuv tsis nkag siab dab tsi liThiab nws tsis yog tias kuv lub neural network muaj teeb meem nrog nws lub mechanism mloog. Kev tsim tag nrho cov architectures no zoo li qee yam hackathon loj heev, qhov twg cov kws sau ntawv sib tw hauv hacks. Hack yog ib qho kev daws teeb meem sai rau qhov teeb meem programming nyuaj. Ntawd yog, tsis muaj kev sib txuas lus pom lossis nkag siab ntawm tag nrho cov architectures no. Txhua yam uas koom ua ke lawv yog ib pawg ntawm cov hacks ua tiav tshaj plaws uas lawv qiv los ntawm ib leeg, ntxiv rau lub hom phiaj sib xws rau lawv txhua tus. kev ua haujlwm ntawm kev tawm tswv yim convolution (rov qab nthuav tawm ntawm qhov yuam kev). Tsis muaj kev xav txog lub cevNws tsis meej tias yuav tsum hloov dab tsi lossis yuav ua li cas thiaj ua kom cov kev ua tiav uas twb muaj lawm zoo dua.

Vim tias cov hacks tsis muaj kev sib txuas lus zoo, lawv nyuaj heev rau nco qab thiab siv. Lawv cov kev paub yog fragmented. Qhov zoo tshaj plaws, ob peb lub ntsiab lus nthuav thiab tsis tau xav txog tau nco qab, tab sis feem ntau ntawm qhov tau nkag siab thiab qhov tsis tau ploj mus hauv ob peb hnub. Nws yuav zoo yog tias koj tuaj yeem nco qab lub npe ntawm lub tsev tom qab ib lub lim tiam. Thiab tseem, ntau teev thiab txawm tias hnub ntawm kev ua haujlwm tau nkim sijhawm nyeem cov ntawv xov xwm thiab saib cov yeeb yaj kiab piav qhia!

Daim Duab 2 – Lub vaj tsiaj 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 kom ntseeg tau tias txawm tias qhov kev paub tawg no tsis nkag siab los ntawm tus nyeem ntawv. Tab sis cov kab lus adverbial participial hauv kaum kab lus nrog cov qauv rub tawm ntawm huab cua nyias nyias yog lub ncauj lus rau ib tsab xov xwm sib cais (teeb meem tshaj tawm lossis piam sij).

Vim li no, muaj qhov xav tau los tsim cov ntaub ntawv ntawm cov neural network thiab, yog li, txhim kho qhov zoo ntawm kev nkag siab thiab kev nco qab. Yog li ntawd, lub ntsiab lus tseem ceeb ntawm kev tshuaj xyuas cov thev naus laus zis thiab cov qauv ntawm cov neural network dag tau dhau los ua txoj haujlwm hauv qab no: nrhiav seb qhov no mus qhov twg, thiab tsis yog cov qauv ntawm ib qho neural network tshwj xeeb hauv kev sib cais.

Qhov no mus qhov twg lawm? Cov ntsiab lus tseem ceeb:

• Tus naj npawb ntawm cov kev pib kawm tshuab hauv ob xyoos dhau los poob qis heevTej zaum yog vim li cas: "cov neural networks tsis yog ib yam tshiab lawm."
• Txhua tus neeg tuaj yeem tsim ib lub neural network ua haujlwm los daws teeb meem yooj yim. Txhawm rau ua qhov no, coj tus qauv npaj txhij los ntawm "qauv vaj tsiaj" thiab cob qhia txheej kawg ntawm lub neural network (hloov kev kawm) ntawm cov ntaub ntawv npaj txhij los ntawm Google Dataset Nrhiav lossis ntawm 25 cov ntaub ntawv Kaggle nyob rau hauv dawb Jupyter Notebook huab.
• Cov chaw tsim khoom loj ntawm neural network tau pib tsim "qauv zoos" (qauv vaj tsiaj). Nrog lawv txoj kev pab, koj tuaj yeem tsim daim ntawv thov lag luam sai sai: TF Hub rau TensorFlow, Kev Tshawb Pom MM rau PyTorch, Detectron rau Caffe2, chainer-modelzoo rau Chainer thiab lwm yam.
• Cov tes hauj lwm neural ua haujlwm hauv lub sijhawm tiag tiag (lub sijhawm tiag tiag) ntawm cov khoom siv mobile. Txij li 10 txog 50 thav duab ib ob.
• Daim ntawv thov ntawm 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 nqa neural networks ntawm qib kho vajtse (neural accelerators).
• Txhua yam khoom siv, ib daim khaub ncaws, thiab tej zaum txawm tias zaub mov yuav muaj Chaw nyob IP-v6 thiab sib txuas lus nrog ib leeg" - Sebastian Thrun.
• Tus naj npawb ntawm cov ntawv tshaj tawm txog kev kawm tshuab tau pib loj hlob tshaj Moore txoj cai (ob npaug txhua ob xyoos) txij li xyoo 2015. Nws yog qhov tseeb tias, cov neural networks rau kev tshuaj xyuas tsab xov xwm yog qhov xav tau.
• Cov thev naus laus zis hauv qab no tau txais kev nyiam:
  • PyTorch - qhov nrov npe tab tom loj hlob sai thiab zoo li yuav dhau TensorFlow.
  • Kev xaiv hyperparameter tsis siv neeg AutoML - qhov nrov zuj zus zoo.
  • Maj mam txo qhov tseeb thiab nce qhov ceev ntawm kev suav: kev xav tsis meej, cov algorithms txhawb nqa, kev suav tsis meej (kwv yees li), kev suav lej (thaum qhov hnyav ntawm lub neural network raug hloov mus ua cov lej thiab suav lej), neuroaccelerators.
  • Hloov Mus cov duab rau cov ntawv nyeem и ntawv rau duab.
  • creation Cov khoom 3D los ntawm cov yeeb yaj kiab, tam sim no nyob rau hauv lub sijhawm tiag tiag.
  • Tus yuam sij rau DL yog ntau cov ntaub ntawv, tab sis kev sau thiab sau npe rau nws tsis yooj yim. Yog li ntawd, kev sau npe tsis siv neeg tab tom txhim kho (kev sau ntawv tsis siv neeg) rau cov neural networks siv neural networks.
• Nrog rau neural networks, Computer Science mam li nco dheev los ua kev tshawb fawb txog kev sim thiab sawv tsees kev kub ntxhov ntawm kev rov ua dua tshiab.
• Cov nyiaj IT thiab qhov nrov ntawm neural networks tau tshwm sim tib lub sijhawm, thaum kev suav lej tau los ua tus nqi lag luam. Kev lag luam tau hloov pauv los ntawm kev lag luam kub-txiaj. kev suav nyiaj kubSaib kuv tsab xov xwm ntawm kev tshawb fawb txog kev lag luam thiab yog vim li cas rau qhov tshwm sim ntawm IT nyiaj.

Ib qho tshiab maj mam tshwm sim Txoj kev siv ML/DL los tsim cov program (Kev Kawm Tshuab & Kev Kawm Sib Sib Zog), uas yog raws li kev sawv cev rau ib qho kev pab cuam ua ib pawg ntawm cov qauv neural network uas tau kawm tiav.

Daim Duab 3 - ML/DL ua ib txoj kev qhia tshiab txog kev sau program

Txawm li cas los xij, nws yeej tsis tau tshwm sim cov kev tshawb fawb txog neural network, uas ib tug neeg tuaj yeem xav thiab ua haujlwm tsis tu ncua. Qhov uas tam sim no hu ua "kev tshawb fawb" yog qhov tseeb, kev sim, heuristic algorithms.

Cov kev sib txuas rau kuv thiab lwm cov peev txheej:

• Tsab ntawv xov xwm txog Kev Tshawb Fawb Txog Cov Ntaub Ntawv. Feem ntau yog hais txog kev ua cov duab. Cov neeg uas xav tau txais nws tuaj yeem xa email rau kuv (foobar167<gaf-gaf>gmail<dot>com). Kuv yuav xa cov kev sib txuas mus rau cov ntawv xov xwm thiab cov yeeb yaj kiab thaum kuv sau cov ntaub ntawv tshiab.
• Tag Nrho daim ntawv teev cov chav kawm thiab cov ntawv sauuas kuv tau dhau los thiab uas kuv xav dhau.
• Cov chav kawm thiab cov yeeb yaj kiab rau cov neeg pib tshiab, uas tsim nyog pib kawm txog neural networks. Ntxiv rau ib daim ntawv qhia Kev Taw Qhia Txog Kev Kawm Tshuab thiab Kev Siv Tshuab Ua Haujlwm.
• Cov cuab yeej siv tau, qhov twg txhua tus neeg yuav pom qee yam nthuav rau lawv tus kheej.
• Lawv tau hloov tawm tias muaj txiaj ntsig zoo heev cov channel video rau kev tshuaj xyuas cov ntawv tshawb fawb Kev Tshawb Fawb Txog Cov Ntaub Ntawv. Nrhiav lawv, ua raws li lawv, thiab qhia cov kev sib txuas rau koj cov npoj yaig thiab kuv. Piv txwv li:
  • Ob Foob Cov Ntawv
  • Henry AI Labs
  • Yannic Kilcher
  • CodeEmporium
  • Blog Chengwei Zhang aka Tony 607 nrog cov lus qhia ib kauj ruam zuj zus thiab cov cai qhib qhov chaw.

Ua tsaug rau koj txoj kev paub!

Tau qhov twg los: www.hab.com