Mamodheru eKudzidza eMichina ari kwese kwese iko zvino. Pakati pezuva, iwe unogona kushandisa aya mamodheru zvakanyanya kupfuura zvaunofunga. Mamodheru ekudzidza emuchina anoshandiswa mumabasa akajairwa akadai sekubhurawuza kwesocial media, kutora mafoto, uye kutarisa mamiriro ekunze.
Muchina-kudzidza algorithm inogona kunge yakakurudzira iyi blog kwauri. Tese takanzwa nezvekuti zvinotora nguva sei kudzidzisa mamodheru aya. Tese takanzwa kuti kudzidzisa mamodheru aya kunodya nguva.
Nekudaro, kuita fungidziro pamhando idzi kunowanzo kudhura computational.
Tinoda masisitimu emakombuta anokurumidza kubata chiyero chatiri kushandisa masevhisi ekudzidza muchina. Nekuda kweizvozvo, mazhinji emamodheru aya anomhanyiswa panzvimbo huru dzedata neCPU neGPU masumbu (kunyangwe TPU mune dzimwe nguva).
Paunotora mufananidzo, unoda machine learning kuti uvandudze pakarepo. Iwe haudi kumirira kuti mufananidzo uendeswe kunzvimbo yedata, ugadziriswe, uye udzoserwe kwauri. Muchiitiko ichi, modhi yekudzidza yemuchina inofanirwa kuitwa munharaunda.
Paunoti "Hesi Siri" kana "Zvakanaka, Google," unoda kuti zvishandiso zvako zvipindure nekukasika. Kumirira kuti izwi rako riendeswe kumakomputa, kwarinozoongororwa uye data riwanikwe.
Izvi zvinotora nguva uye zvine mhedzisiro yakaipa pane ruzivo rwemushandisi. Muchiitiko ichi, iwe unoda kuti modhi yekudzidza yemuchina ishande munharaunda zvakare. Apa ndipo panopinda TinyML.
Mune ino post, isu tichatarisa muTinyML, kuti inoshanda sei, mashandisiro ayo, maitiro ekutanga nayo, nezvimwe zvakawanda.
Chii TinyML?
TinyML chirango chekucheka-kumucheto icho chinoshandisa shanduko yekudzidza kwemuchina pakuita uye masimba emagetsi ezvishandiso zvidiki uye masisitimu akaiswa.
Kubudirira kuendesa muindasitiri iyi kunoda kunyatsonzwisiswa kwezvishandiso, algorithms, Hardware, uye software. Muchina wekudzidza subgenre inoshandisa yakadzama kudzidza uye muchina kudzidza modhi mumasisitimu akaiswa anoshandisa mamicrocontrollers, madhijitari masaini processors, kana mamwe ma-ultra-low-power specialized processors.
TinyML-inogonesa yakamisikidzwa michina inoitirwa kumhanyisa muchina kudzidza algorithm yerimwe basa, kazhinji sechikamu chemudziyo wacho. edge computing.
Kuti umhanye kwemavhiki, mwedzi, kana makore usingachajizve kana kutsiva bhatiri, masisitimu aya akaiswa anofanira kunge aine simba rekushandisa risingasviki 1 mW.
Sei kushanda?
Iyo yega yemuchina yekudzidza sisitimu inogona kushandiswa ine microcontrollers uye makomputa ndeye TensorFlow Lite. Iyo seti yezvishandiso zvinoita kuti vanogadzira vamhanye mamodheru avo pane nharembozha, yakamisikidzwa, uye yemupendero michina, ichibvumira kudzidza kwemichina panhunzi.
Iyo microcontroller interface inoshandiswa kuunganidza data kubva kumasensa (senge maikorofoni, makamera, kana ma sensors akaiswa mukati).
Isati yatumirwa kune microcontroller, iyo data inosanganiswa mune yegore-yakavakirwa muchina yekudzidza modhi. Batch kudzidziswa muoffline mode inowanzo shandiswa kudzidzisa aya mamodheru. Iyo sensor data ichashandiswa kudzidza uye kufungidzira yakatotemerwa kushandiswa chaiko.
Kana iyo modhi iri kudzidziswa kuona izwi rekumuka, semuenzaniso, yakatomisikidzwa kuti ibate inoenderera inoyerera inoyerera kubva maikorofoni.
Zvese zvakatoitwa nerubatsiro rwepuratifomu yegore senge Google Colab munyaya yeTensorFlow Lite, kusanganisira kusarudzwa kwedhatabheti, kugarisa, kusakwana kana kuwandisa kwemuenzaniso, kudzoreredza, kuwedzera data, kudzidziswa, kusimbiswa, uye kuyedzwa.
Modhi yakadzidziswa zvizere inozopedzisira yashandurwa uye yoendeswa kune microcontroller, microcomputer, kana dhijitari chiratidzo processor mushure mekunze batch kudzidziswa. Iyo modhi haina imwe dzidziso mushure mekuendeswa kune yakamisikidzwa mudziyo. Pane kudaro, inongoshandisa chaiyo-nguva data kubva kumasensa kana michina yekupinza kushandisa modhi.
Nekuda kweizvozvo, TinyML muchina wekudzidza modhi unofanirwa kuve wakasimba zvakanyanya uye uchikwanisa kudzidziswazve mushure memakore kana kusambodzidziswazve. Yese inogoneka modhi yakaderera uye yakawandisa inofanirwa kuongororwa kuitira kuti modhi irambe yakakosha kwenguva yakareba, zvisingagumi.
Asi Sei uchishandisa TinyML?
TinyML yakatanga sekuyedza kubvisa kana kuderedza kuvimba kweIoT pamasevhisi egore kune diki-diki. machine learning mabasa. Izvi zvaida kushandiswa kwemichina yekudzidza modhi pamidziyo yemupendero pachayo. Inopa zvinotevera zvakakosha zvinobatsira:
- Yakaderera-simba kunwa: A TinyML application inofanirwa kushandisa isingasviki 1 milliWatt yesimba. Nekushandisa kwakaderera-simba kwakadaro, mudziyo unogona kuramba uchiwana mhedziso kubva kune sensor data kwemwedzi kana makore, kunyangwe ichifambiswa nebhatiri remari.
- Mari shoma: Yakagadzirirwa kumhanya pane yakaderera-mutengo 32-bit microcontrollers kana DSPs. Aya ma microcontrollers anowanzo masendi mashoma ega ega, uye yakazara yakamisikidzwa sisitimu yakagadziridzwa nawo isingasviki madhora makumi mashanu. Iyi isarudzo inodhura kwazvo yekumhanyisa zvirongwa zvidiki zvekudzidza muchina pachiyero chikuru, uye inonyanya kubatsira muIoT maapplication uko kunofanirwa kushandiswa muchina kudzidza.
- Lower Latency: Zvishandiso zvaro zvine low latency sezvo zvisingade kutakura kana kuchinjana data netiweki. Yese sensor data inorekodhwa munharaunda, uye mhedziso dzinotorwa uchishandisa modhi yakatodzidziswa. Mhedzisiro yekufungidzira inogona kutumirwa kuseva kana gore rekutema matanda kana kuwedzera kugadzirisa, kunyangwe izvi zvisina kukosha kuti mudziyo ushande. Izvi zvinodzikisira network latency uye zvinobvisa kudiwa kwemashini ekudzidzira maoperation kuti aitwe pagore kana server.
- Privacy: Icho chinonetsa zvakanyanya painternet uye neinternet yezvinhu. Basa rekudzidza muchina muTinyML rinoitwa munharaunda, pasina kuchengetedza kana kutumira sensor / mushandisi data kune server/gore. Nekuda kweizvozvo, kunyangwe zvakabatana netiweki, izvi zvikumbiro zvakachengeteka kushandisa uye hazvina njodzi yekuvanzika.
Applications
- Kurima – Rini varimi vanotora mufananidzo wechirimwa, TensorFlow Lite's application inoona zvirwere mairi. Inoshanda pane chero mudziyo uye haidi chinongedzo cheinternet. Nzira iyi inochengetedza zvido zvekurima uye chinhu chakakosha kuvarimi vekumaruwa.
- Mechanics Maintenance -TinyML, kana yakashandiswa pamidziyo ine simba shoma, inogona kuramba ichiona kukanganisa mumuchina. Inosanganisira kufanotaura-kwakavakirwa kuchengetedza. Ping Services, yekutanga-yeAustralia, yakaunza gajeti yeIoT inotarisisa maturbine emhepo nekuzvisungirira kune turbine kunze. Inozivisa zviremera pese painowona chero dambudziko ringaite kana kusashanda zvakanaka.
- Zvipatara – The Solar Scare chirongwa. Umhutu hunoshandisa TinyML kumisa kupararira kwezvirwere zvakaita sedengue nemarariya. Inofambiswa nesimba rezuva uye inoona mamiriro ekuberekesa umhutu isati yaratidza mvura kuti inodzivisa kubereka kweumhutu.
- Traffic Surveillance – By kushandisa TinyML kumasensa anounganidza chaiyo-nguva traffic data, tinogona kudzishandisa kunanga traffic uye kucheka nguva dzekupindura dzemotokari dzenjodzi. Swim.AI, semuenzaniso, inoshandisa tekinoroji iyi pakutepfenyura data kuwedzera kuchengetedzeka kwevatyairi uku ichidzikisawo kuungana uye kuburitswa kwemhepo kuburikidza nekuchenjera nzira.
- mutemo: TinyML inogona kushandiswa mumutemo kuona zviito zvisiri pamutemo zvakaita sebongozozo uye kuba uchishandisa muchina kudzidza uye kuzivikanwa nemasaini. Chirongwa chakafanana chinogona kushandiswawo kuchengetedza maATM ebhangi. Nekuona maitiro emushandisi, modhi yeTinyML inogona kufanotaura kana mushandisi ari mutengi chaiye ari kupedzisa kutengeserana kana muparidzi ari kuedza kubaya kana kuparadza ATM.
Ungatanga sei neTinyML?
Kuti utange neTinyML muTensorFlow Lite, iwe unozoda inoenderana microcontroller board. TensorFlow Lite yeMicrocontrollers inotsigira mamicrocontrollers akanyorwa pazasi.
- Wio Terminal: ATSAMD51
- Himax WE-I Plus EVB Endpoint AI Development Board
- STM32F746 Discovery kit
- Adafruit EdgeBadge
- Synopsys DesignWare ARC EM Software Development Platform
- Sony Express
- Arduino Nano 33 BLE Sense
- SparkFun Edge
- Adafruit TensorFlow Lite yeMicrocontrollers Kit
- Adafruit Circuit Playground Bluefruit
- Espressif ESP32-DevKitC
- Espressif ESP-EYE
Aya ndiwo 32-bit microcontrollers ane akakwana flash memory, RAM, uye wachi frequency yekuitisa muchina kudzidza modhi. Iwo mabhodhi zvakare ane akati wandei epaboard sensors anokwanisa kumhanyisa chero chirongwa chakamisikidzwa uye kushandisa michina yekudzidza modhi kune yakanangwa application. To gadzira modhi yekudzidza muchina, iwe uchada laptop kana kombiyuta kuwedzera kune hardware platform.
Imwe neimwe puratifomu yehardware ine yayo yekuronga maturusi ekuvaka, kudzidzisa, uye porting muchina kudzidza modhi, iyo inoshandisa TensorFlow Lite yeMicrocontrollers package. TensorFlow Lite ndeyemahara kushandisa uye kugadzirisa nekuti ndizvo open source.
Kuti utange neTinyML neTensorFlow Lite, zvese zvaunoda ndeimwe yeakataurwa pamusoro akamisikidzwa mapuratifomu ehardware, komputa/laptop, USB tambo, USB-to-Serial inoshandura - uye chishuwo chekudzidzira muchina kudzidza nemasisitimu akaiswa. .
matambudziko
Kunyangwe kufambira mberi kweTinyML kwaunza mhedzisiro yakanaka, indasitiri yekudzidza yemuchina ichiri kusangana nezvipingaidzo zvakakura.
- Kusiyanisa kweSoftware - Hand-coding, kugadzirwa kwekodhi, uye vaturikiri veML zvese zvingasarudzwa zvekutumira modhi paTinyML zvishandiso, uye imwe neimwe inotora huwandu hwakasiyana hwenguva nesimba. Kuita kwakasiyana kunogona kuitika nekuda kweizvi.
- Hardware kusiyana - Ikoko kune akati wandei Hardware sarudzo dziripo. TinyML mapuratifomu anogona kuve chero chinhu kubva kune-yakawanda-chinangwa microcontrollers kune yekucheka-kumucheto neural processors. Izvi zvinokonzeresa nyaya nekutumirwa kwemodhi pazvivakwa zvakasiyana.
- Kugadzirisa / kugadzirisa - Nguva iyo ML modhi inoita zvisina kunaka pagore, zviri nyore kutarisa data uye kuona kuti chii chiri kunetsa. Kana modhi yakapararira muzviuru zveTinyML zvishandiso, pasina dhata rinodzokera kugore, kugadzirisa dambudziko kunonetsa uye kungangoda imwe nzira.
- Memory constraints - Traditional mapuratifomu, akadai semafoni nemalaptops, anoda gigabytes ye RAM, nepo TinyML zvishandiso zvinoshandisa kilobytes kana megabytes. Nekuda kweizvozvo, ukuru hweiyo modhi inogona kutumirwa ishoma.
- Model kudzidziswa - kunyangwe pane zvakati wandei zvakanakira kuendesa ML modhi paTinyML zvishandiso, iyo yakawanda yeML modhi ichiri kudzidziswa pamusoro pegore kuti iite iterate uye kuenderera mberi nekuvandudza mamodhi.
Future
TinyML, ine diki tsoka, yakaderera bhatiri kushandiswa, uye kushomeka kana kushomeka kuvimba neinternet yekubatanidza, ine hukuru hwekugona mune ramangwana, sezvo ruzhinji rwakamanikana. chakagadzirwa njere ichaitwa pamidziyo yemupendero kana yakazvimirira yakamisikidzwa gadget.
Ichaita kuti maIoT maapplication awedzere kuvanzika uye akachengeteka nekuvasimudzira. Nyangwe TensorFlow Lite parizvino ndiyo yega sisitimu yekudzidza yemuchina yemamicrocontrollers uye mamicrocomputer, mamwe maficha akafanana sesensor uye ARM's CMSIS-NN ari mumabasa.
Nepo TensorFlow Lite iri yakavhurika-sosi purojekiti iri kuenderera mberi iyo yakatanga kutyisa neChikwata cheGoogle, ichiri kuda rutsigiro rwenharaunda kuti ipinde mune mainstream.
mhedziso
TinyML inzira itsva inosanganisa masisitimu akadzamirirwa nekudzidza kwemichina. Sezvo iyo yakatetepa AI inokwira mune akawanda verticals uye madomasi, iyo tekinoroji inogona kubuda seyakakurumbira subfield mukudzidza kwemichina uye nehungwaru hwekugadzira.
Inopa mhinduro kune akawanda matambudziko ayo iyo IoT chikamu uye nyanzvi dzinoshandisa muchina kudzidza kune akawanda madomasi-chaiwo maitiro akatarisana.
Pfungwa yekushandisa muchina kudzidza pa midziyo yemupendero ine diki komputa footprint uye simba rekushandisa rine mukana wekushandura zvakanyanya magadzirirwo akaiswa masisitimu uye marobhoti.
Leave a Reply