Iimodeli zokuFunda ngoomatshini zikuyo yonke indawo ngoku. Emini, mhlawumbi usebenzisa ezi modeli kakhulu kunokuba ucinga. Iimodeli zokufunda ngoomatshini zisetyenziswa kwimisebenzi eqhelekileyo efana nokukhangela kwimidiya yoluntu, ukufota, kunye nokujonga imozulu.
I-algorithm yokufunda ngomatshini inokuba ikuncomele le bhlog. Sonke sivile malunga nokuba kuthatha ixesha ukuqeqesha ezi modeli. Sonke sivile ukuba ukuqeqesha ezi modeli kuthatha ixesha.
Nangona kunjalo, ukwenza intelekelelo kwezi modeli kuhlala kubiza kakhulu.
Sifuna iinkqubo zekhompyuter ezikhawulezayo ukumelana nezinga esisebenzisa ngalo iinkonzo zokufunda ngoomatshini. Ngenxa yoko, uninzi lwezi modeli ziqhutywa kumaziko amakhulu edatha ane-CPU kunye namaqela e-GPU (kunye nee-TPU kwezinye iimeko).
Xa uthatha umfanekiso, ufuna yokufunda umatshini ukuyiphucula kwangoko. Awufuni ukulinda ukuba umfanekiso udluliselwe kwiziko ledatha, iqwalaselwe, kwaye ibuyiselwe kuwe. Kule meko, imodeli yokufunda koomatshini kufuneka iqhutywe ekuhlaleni.
Xa usithi "Hey Siri" okanye "Kulungile, Google," ufuna izixhobo zakho ziphendule ngokukhawuleza. Ilinde ukuba ilizwi lakho lidluliselwe kwiikhompyuter, apho liya kuvavanywa kwaye kufumaneke idatha.
Oku kuthatha ixesha kwaye kunefuthe elibi kumava omsebenzisi. Kule meko, ufuna imodeli yokufunda koomatshini isebenze nalapha ekhaya. Apha kulapho iTinyML ingena khona.
Kule posi, siza kujonga kwi-TinyML, indlela esebenza ngayo, ukusetyenziswa kwayo, ukuqalisa ngayo, kunye nokunye okuninzi.
Yintoni i I-TinyML?
I-TinyML luqeqesho olubukhali olusebenzisa amandla okuguqula umatshini wokufunda ekusebenzeni kunye nemida yamandla yezixhobo ezincinci kunye neenkqubo ezizinzisiweyo.
Ukuthunyelwa ngempumelelo kolu shishino kufuna ukuqondwa ngokucokisekileyo kwezicelo, i-algorithms, ihardware, kunye nesoftware. Luhlelo olusezantsi lokufunda ngomatshini olusebenzisa ukufunda okunzulu kunye neemodeli zokufunda ngoomatshini kwiinkqubo ezizinzisiweyo ezisebenzisa i-microcontrollers, iiprosesa zesignali zedijithali, okanye ezinye iiprosesa ezikhethekileyo zamandla asezantsi.
Izixhobo ezizinzisiweyo ze-TinyML zenzelwe ukuqhuba i-algorithm yokufunda ngomatshini kumsebenzi othile, njengenxalenye yesixhobo. khompyutha.
Ukuze kuqhutywe iiveki, iinyanga, okanye iminyaka ngaphandle kokutshaja ngokutsha okanye ukutshintshwa kwebhetri, ezi sistim zifakelweyo kufuneka zisebenzise amandla angaphantsi kwe-1 mW.
Ingaba isebenza kanjani?
Ekuphela kwesakhelo sokufunda koomatshini esinokusetyenziswa kunye ne-microcontroller kunye neekhompyuter I-TensorFlow Lite. Yiseti yezixhobo ezivumela abaphuhlisi ukuba baqhube iimodeli zabo kwiselfowuni, ezizinzisiweyo, kunye nezixhobo ezisemaphethelweni, ezivumela ukufundwa koomatshini ngokubhabha.
I-interface ye-microcontroller isetyenziselwa ukuqokelela idatha kwi-sensors (njenge-microphone, iikhamera, okanye i-sensor ezifakwe ngaphakathi).
Ngaphambi kokuba ithunyelwe kwi-microcontroller, idatha ifakwe kwimodeli yokufunda yomatshini esekelwe kwifu. Uqeqesho lwebhetshi kwimowudi engasebenziyo idla ngokusetyenziswa ukuqeqesha ezi modeli. Idata yoluvo oluya kusetyenziselwa ukufunda kunye nokuqonda sele imiselwe kwisicelo esithile.
Ukuba imodeli iqeqeshelwa ukufumanisa igama lokuvuka, umzekelo, sele isetyenzisiwe ukuphatha i-audio stream eqhubekayo evela kwi-microphone.
Yonke into sele yenziwe ngoncedo lweqonga lelifu elifana neGoogle Colab kwimeko yeTensorFlow Lite, kubandakanywa ukukhethwa kwedatha, ukulungelelanisa, ukufaneleka okanye ukugqithisa imodeli, ukulungelelanisa, ukwandiswa kwedatha, uqeqesho, ukuqinisekiswa, kunye nokuvavanya.
Imodeli eqeqeshwe ngokupheleleyo ekugqibeleni iguqulwa kwaye idluliselwe kwi-microcontroller, i-microcomputer, okanye iprosesa yesignali yedijithali emva koqeqesho lwebhetshi engaxhunyiwe kwi-intanethi. Imodeli ayinalo uqeqesho olongezelelweyo emva kokuba ihanjiswe kwisixhobo esifakwe ngaphakathi. Endaweni yoko, isebenzisa kuphela idatha yexesha lokwenyani ukusuka kuluvo okanye izixhobo zokufaka ukufaka imodeli.
Ngenxa yoko, imodeli yokufunda yomatshini we-TinyML kufuneka yomelele kwaye ikwazi ukuqeqeshwa emva kweminyaka okanye ingaze iphinde iqeqeshwe. Yonke imodeli enokubakho ingaphantsi kunye nokufakwa ngokugqithisileyo kufuneka iphandwe ukuze imodeli ihlale ifanelekile ixesha elide, ngokungenasiphelo.
Kodwa kutheni usebenzisa iTinyML?
I-TinyML yaqala njengenzame zokuphelisa okanye ukunciphisa ukuthembela kwe-IoT kwiinkonzo zelifu kwisiseko esincinci esincinci. yokufunda umatshini imisebenzi. Oku kwenze imfuneko yokusetyenziswa kweemodeli zokufunda koomatshini kwizixhobo ezisekupheleni ngokwazo. Ibonelela ngezi nzuzo zilandelayo:
- Amandla asezantsi ukusebenzisa: Isicelo seTinyML kufuneka sisebenzise ngaphantsi kwe-1 milliWatt yamandla. Ngokusetyenziswa kwamandla aphantsi ngolo hlobo, isixhobo sinokuqhubeka sifumana izigqibo kwidatha yesivamvo kangangeenyanga okanye iminyaka, nokuba inikwe ibhetri yemali.
- Iindleko eziphantsi: Yenzelwe ukusebenza kwii-microcontrollers ze-32-bit okanye i-DSPs. Ezi microcontrollers ziqhelekile ukuba ziisenti ezimbalwa nganye, kwaye iyonke inkqubo edibeneyo ephuhliswe kunye nabo ingaphantsi kwe-50 yeedola. Olu lukhetho oluxabisa kakhulu ukuqhuba iinkqubo zokufunda koomatshini kumlinganiselo omkhulu, kwaye luluncedo ngakumbi kwizicelo ze-IoT apho kufuneka kusetyenziswe umatshini wokufunda.
- Ukubambezeleka okusezantsi: Izicelo zayo zine-latency ephantsi kuba ayifuni ukuthutha okanye ukutshintshiselana ngedatha kwinethiwekhi. Yonke idatha yoluvo irekhodwa kwindawo, kwaye izigqibo zithathwa kusetyenziswa imodeli esele iqeqeshiwe. Iziphumo zentelekelelo zinokuthunyelwa kwiseva okanye ilifu lokuloga okanye ukuqhubekekiswa okongeziweyo, nangona oku kungabalulekanga ukuba isixhobo sisebenze. Oku kunciphisa i-latency yothungelwano kwaye kuphelisa isidingo sokusebenza komatshini wokufunda okwenziwa kwilifu okanye kwiseva.
- Ukuba bucala: Yinkxalabo enkulu kwi-intanethi kunye ne-intanethi yezinto. Umsebenzi wokufunda koomatshini kwii-apps ze-TinyML wenziwa ekuhlaleni, ngaphandle kokugcina okanye ukuthumela i-sensor/idatha yomsebenzisi kwiseva/kwilifu. Ngenxa yoko, nangona ziqhagamshelwe kwinethiwekhi, ezi zicelo zikhuselekile ukuba zingasetyenziswa kwaye azikho mngcipheko wabucala.
izicelo
- EzoLimo – Nini amafama athatha ifoto yesityalo, isicelo sikaTensorFlow Lite sibona izifo kuyo. Isebenza kuso nasiphi na isixhobo kwaye ayifuni uxhumano lwe-intanethi. Inkqubo ikhusela umdla wezolimo kwaye iyimfuneko ebalulekileyo kumafama asemaphandleni.
- ULondolozo lweeMechanics -I-TinyML, xa isetyenziswe kwizixhobo ezinamandla aphantsi, inokuqhubeka ibona iimpazamo kumatshini. Ibandakanya ugcino olusekelwe kuqikelelo. Iinkonzo ze-Ping, isiqalo sase-Australia, sazise igajethi ye-IoT ebeka iliso kwiiinjini zomoya ngokuzincamathela ngaphandle kwe-injini yomoya. Yazisa abasemagunyeni nanini na xa ifumanisa nayiphi na ingxaki enokwenzeka okanye ukungasebenzi kakuhle.
- Izibhedlele – The ISolar Scare yiprojekthi. Iingcongconi zisebenzisa i-TinyML ukunqanda ukusasazeka kwezifo ezifana nedengue kunye nemalariya. Isebenza ngamandla elanga kwaye ibona iimeko zokuzala kweengcongconi phambi kokuba inike umqondiso wokuba amanzi anqande ukuzala kweengcongconi.
- Ukujongwa kweTrafikhi – Ngu ukusebenzisa i-TinyML kwiinzwa eziqokelela idatha ye-traffic yexesha langempela, sinokuzisebenzisa ukuqondisa ngcono i-traffic kunye nokusika amaxesha okuphendula kwizithuthi eziphuthumayo. I-Swim.AI, umzekelo, isebenzisa le teknoloji kwidatha yokusasaza ukunyusa ukhuseleko lwabakhweli ngelixa ikwanciphisa ukuxinana kunye nokukhutshwa kwe-smart routing.
- umthetho: I-TinyML ingasetyenziswa kunyanzeliso lomthetho ukuchonga izenzo ezingekho mthethweni ezinje ngoqhushululu kunye nobusela kusetyenziswa umatshini wokufunda kunye nokuqaphela izimbo zomzimba. Inkqubo efanayo isenokusetyenziswa ukukhusela ii-ATM zebhanki. Ngokujonga ukuziphatha komsebenzisi, imodeli ye-TinyML inokuqikelela ukuba ngaba umsebenzisi ngumthengi wokwenyani ogqibezela intengiselwano okanye umngeneleli ozama ukugqekeza okanye ukutshabalalisa i-ATM.
Ungaqalisa njani ngeTinyML?
Ukuze uqalise nge-TinyML kwi-TensorFlow Lite, uya kudinga ibhodi ye-microcontroller ehambelanayo. I-TensorFlow Lite yeMicrocontrollers ixhasa ii-microcontrollers ezidweliswe ngezantsi.
- Itheminali yeWio: ATSAMD51
- Ibhodi yoPhuhliso lwe-Himax WE-I Plus EVB Endpoint AI
- STM32F746 ikhithi yokuFumana
- Adafruit EdgeBadge
- I-Synopsys DesignWare ARC EM iQonga loPhuhliso lweSoftwe
- I-Sony Spresense
- UArduino Nano 33 BLE Sense
- I-SparkFun Edge
- I-Adafruit TensorFlow Lite yeMicrocontrollers Kit
- Adafruit Circuit Playground Bluefruit
- I-Espressif ESP32-DevKitC
- Espressif ESP-EYE
Ezi zii-32-bit microcontrollers ezinememori eyaneleyo ye-flash, i-RAM, kunye ne-clock frequency ukwenza imodeli yokufunda yomatshini. Iibhodi nazo zinenani leenzwa zebhodi ezikwazi ukuqhuba nayiphi na inkqubo edibeneyo kunye nokusebenzisa iimodeli zokufunda ngomatshini kwisicelo esijoliswe kuyo. Ukuya ukwakha imodeli yokufunda koomatshini, uya kufuna ilaptop okanye ikhompyutha ukongeza kwiqonga lehardware.
Iqonga ngalinye le-hardware linezixhobo zalo zokucwangcisa zokwakha, uqeqesho, kunye neemodeli zokufunda zoomatshini, esebenzisa iTensorFlow Lite yeMicrocontrollers package. I-TensorFlow Lite isimahla ukuyisebenzisa kwaye uyiguqule kuba injalo Vula Umnikezi.
Ukuqalisa nge-TinyML kunye ne-TensorFlow Lite, konke okudingayo yenye yeeplatifti ze-hardware ezichazwe ngasentla, ikhompyutha / i-laptop, intambo ye-USB, i-USB-to-Serial converter - kunye nomnqweno wokuziqhelanisa nokufundwa komatshini kunye neenkqubo ezifakwe ngaphakathi. .
mngeni
Nangona inkqubela ye-TinyML ivelise iziphumo ezininzi ezilungileyo, ishishini lokufunda ngoomatshini lisajongene nemiqobo emikhulu.
- Iyantlukwano yesoftware-Ukufakwa kweekhowudi ngesandla, ukuveliswa kwekhowudi, kunye neetoliki ze-ML zizo zonke iinketho zokuthumela iimodeli kwizixhobo ze-TinyML, kwaye nganye ithatha inani elahlukileyo lexesha kunye nomgudu. Imisebenzi eyahlukeneyo inokuvela ngenxa yoku.
- Ukwahluka kwe-Hardware - Kukho kukho iinketho ezininzi zehardware ezikhoyo. Iiplatifti ze-TinyML zinokuba yiyo nantoni na ukusuka kwii-microcontrollers zenjongo ngokubanzi ukuya kwiiprosesa ze-neural zokusika. Oku kubangela imiba ngonikezelo lwemodeli kuzo zonke izakhiwo ezahlukeneyo.
- Ukulungisa ingxaki/ukulungisa – Nini imodeli ye-ML ayisebenzi kakuhle kwilifu, kulula ukujonga idatha kwaye ufumanise ukuba kuqhubeka ntoni. Xa imodeli isasazwe kumawaka ezixhobo ze-TinyML, kungekho lwazi lubuyela kwilifu, ukulungisa ingxaki kuba nzima kwaye kungafuna indlela eyahlukileyo.
- Imiqobo yememori - Yemveli iiplatifti, ezinjengee-smartphones kunye neelaptops, zifuna iigigabhayithi ze-RAM, kanti izixhobo zeTinyML zisebenzisa iikhilobhayithi okanye iimegabhayithi. Ngenxa yoko, ubungakanani bemodeli enokuthi isetyenziswe bulinganiselwe.
- Uqeqesho olungumzekelo-Nangona kukho iingenelo ezininzi zokusebenzisa imifuziselo ye-ML kwizixhobo ze-TinyML, ubuninzi beemodeli ze-ML zisaqeqeshwa kwilifu ukuze ziphindaphinde kwaye ngokuqhubekayo ziphucule imodeli echanekileyo.
Future
I-TinyML, eneenyawo zayo ezincinci, ukusetyenziswa kwebhetri okuphantsi, kunye nokunqongophala okanye ukuthembela okulinganiselweyo kuqhagamshelwano lwe-intanethi, inamandla amakhulu kwixesha elizayo, njengoko uninzi lwamancinci. kukubhadla okungeyonyani iya kuphunyezwa kwizixhobo zomda okanye izixhobo ezizinzisiweyo ezizimeleyo.
Iyakwenza izicelo ze-IoT zibe yimfihlo kwaye zikhuseleke ngokuzixhasa. Nangona TensorFlow I-Lite okwangoku isesona sikhokelo sokufunda umatshini we-microcontrollers kunye nee-microcomputers, ezinye izikhokelo ezinokuthelekiswa ezifana ne-sensor kunye ne-ARM's CMSIS-NN zisemisebenzini.
Ngelixa i-TensorFlow Lite iyiprojekthi yomthombo ovulekileyo eqhubekekayo eqale ngeQela likaGoogle, isafuna inkxaso yoluntu ukuze ingene kwindawo eqhelekileyo.
isiphelo
I-TinyML yindlela yenoveli edibanisa iinkqubo ezizinzisiweyo kunye nokufunda koomatshini. Njengoko i-AI emxinwa iphakama kwiindawo ezininzi ezithe nkqo kunye nemimandla, itekhnoloji inokuvela njengendawo esezantsi yokufunda koomatshini kunye nobukrelekrele bokwenziwa.
Ibonelela ngesisombululo kwimingeni emininzi icandelo le-IoT kunye neengcali zisebenzisa umatshini wokufunda kwiinkalo ezininzi ezijongene nesizinda esijongene nazo ngoku.
Ingqikelelo yokusebenzisa umatshini wokufunda kwi izixhobo edge nge computing encinane unyawo kunye nokusetyenziswa kwamandla kunamandla okuguqula ngokubonakalayo indlela iinkqubo ezizinzisiweyo kunye neerobhothi zakhiwe.
Shiya iMpendulo