Amamodeli okufunda ngomshini agcwele yonke indawo njengamanje. Emini, cishe usebenzisa lawa mamodeli kakhulu kunalokho okubonayo. Amamodeli okufunda ngomshini asetshenziswa emisebenzini evamile njengokuphequlula inkundla yezokuxhumana, ukuthwebula izithombe, nokuhlola isimo sezulu.
I-algorithm yokufunda ngomshini kungenzeka ikuncomele le bhulogi. Sonke sizwile ukuthi kudla isikhathi kanjani ukuqeqesha lawa mamodeli. Sonke sizwile ukuthi ukuqeqesha lawa mamodeli kudla isikhathi.
Kodwa-ke, ukwenza ukucatshangelwa kulawa mamodeli kuvame ukubiza kakhulu.
Sidinga amasistimu ekhompyutha ashesha ngokwanele ukuphatha izinga esisebenzisa ngalo izinsiza zokufunda zomshini. Ngenxa yalokho, iningi lalawa mamodeli asetshenziswa ezikhungweni zedatha ezinkulu ezinamaqoqo e-CPU ne-GPU (ngisho nama-TPU kwezinye izimo).
Uma uthatha isithombe, ufuna ukufunda imishini ukuyithuthukisa ngokushesha. Awufuni ukulinda ukuthi isithombe sidluliselwe esikhungweni sedatha, sicutshungulwe, futhi sibuyiselwe kuwe. Kulokhu, imodeli yokufunda yomshini kufanele yenziwe endaweni.
Uma uthi “Hey Siri” noma “OK, Google,” ufuna amagajethi akho aphendule ngokushesha. Ilinde ukuthi izwi lakho lidluliselwe kumakhompyutha, lapho lizohlolwa futhi kutholwe nedatha.
Lokhu kuthatha isikhathi futhi kunomthelela omubi kulwazi lomsebenzisi. Kulokhu, ufuna imodeli yokufunda yomshini isebenze nasendaweni. Yilapho i-TinyML ingena khona.
Kulokhu okuthunyelwe, sizobheka i-TinyML, ukuthi isebenza kanjani, ukusetshenziswa kwayo, ukuthi ungaqala kanjani ngayo, nokunye okuningi.
Kuyini I-TinyML?
I-TinyML isiyalo esisezingeni eliphezulu esisebenzisa amandla okuguqula okufunda komshini ekusebenzeni kanye nemikhawulo yamandla yamadivayisi amancane namasistimu ashumekiwe.
Ukuthunyelwa ngempumelelo kulo mkhakha kudinga ukuqonda okuphelele kwezinhlelo zokusebenza, ama-algorithms, ihadiwe, nesoftware. Iwuhlobo olungaphansi lokufunda komshini olusebenzisa ukufunda okujulile namamodeli okufunda omshini kumasistimu ashumekiwe asebenzisa izilawuli ezincane, amaphrosesa esignali yedijithali, noma amanye amaphrosesa akhethekile anamandla aphansi.
Amadivayisi ashumekiwe anikwe amandla i-TinyML ahloselwe ukusebenzisa i-algorithm yokufunda komshini ngomsebenzi othile, ngokuvamile njengengxenye yedivayisi. ikhomputha yekhompyutha.
Ukuze kusebenze amasonto, izinyanga, noma ngisho neminyaka ngaphandle kokushajwa kabusha noma ukushintshwa kwebhethri, lawa masistimu ashumekiwe kufanele abe nokusetshenziswa kwamandla okungaphansi kuka-1 mW.
Isebenza kanjani?
Okuwukuphela kohlaka lokufunda komshini olungasetshenziswa ngama-microcontroller namakhompyutha ngu I-TensorFlow Lite. Kuyisethi yamathuluzi evumela onjiniyela ukuthi basebenzise amamodeli abo kumadivayisi eselula, ashumekiwe, nasemaphethelweni, okuvumela umshini wokufunda ngokushesha.
I-interface ye-microcontroller isetshenziselwa ukuqoqa idatha kusuka kuzinzwa (njengamakrofoni, amakhamera, noma izinzwa ezishumekiwe).
Ngaphambi kokuthunyelwa kusilawuli esincane, idatha ifakwa kumodeli yokufunda yomshini esekelwe emafini. Ukuqeqeshwa kwenqwaba kumodi engaxhunyiwe ku-inthanethi ngokuvamile kusetshenziswa ukuqeqesha lawa mamodeli. Idatha yenzwa ezosetshenziselwa ukufunda kanye nencazelo isivele inqunyelwe isicelo esithile.
Uma imodeli iqeqeshelwa ukuthola igama lokuvuka, isibonelo, isivele isethelwe ukuphatha ukusakaza komsindo okuqhubekayo kusuka kumakrofoni.
Konke sekuvele kwenziwa ngosizo lwenkundla yefu efana ne-Google Colab endabeni ye-TensorFlow Lite, okuhlanganisa ukukhethwa kwedathasethi, ukujwayela, ukufakwa ngaphansi noma ukufakwa ngokweqile kwemodeli, ukujwayela, ukukhulisa idatha, ukuqeqeshwa, ukuqinisekiswa, nokuhlola.
Imodeli eqeqeshwe ngokugcwele ekugcineni iguqulwa futhi idluliselwe ku-microcontroller, i-microcomputer, noma iphrosesa yesignali yedijithali ngemva kokuqeqeshwa kwenqwaba engaxhunyiwe ku-inthanethi. Imodeli ayinakho ukuqeqeshwa okwengeziwe ngemva kokuhanjiswa kudivayisi eshumekiwe. Kunalokho, isebenzisa kuphela idatha yesikhathi sangempela evela kuzinzwa noma kumadivayisi okokufaka ukuze kusetshenziswe imodeli.
Ngenxa yalokho, imodeli yokufunda yomshini we-TinyML kufanele iqine ngendlela emangalisayo futhi ikwazi ukuqeqeshwa kabusha ngemva kweminyaka noma ingalokothi iqeqeshwe kabusha. Wonke amamodeli okungenzeka afaneleka ngaphansi kanye nokufakwa ngokweqile kufanele kuphenywe ukuze imodeli ihlale ifaneleka isikhathi eside, ngokungenasiphelo.
Kodwa Kungani usebenzisa i-TinyML?
I-TinyML iqale njengomzamo wokuqeda noma ukunciphisa ukuthembela kwe-IoT kumasevisi wamafu ngezinga elincane eliyisisekelo. ukufunda imishini imisebenzi. Lokhu kwenze kudingeke ukuthi kusetshenziswe amamodeli okufunda omshini kumadivayisi onqenqema ngokwawo. Inikeza izinzuzo ezinkulu ezilandelayo:
- Amandla aphansi ukusetshenziswa: Uhlelo lokusebenza lwe-TinyML kufanele lusebenzise amandla angaphansi kwe-1 milliWatt. Ngokusetshenziswa kwamandla aphansi kangako, idivayisi ingase iqhubeke nokuthola iziphetho kusukela kudatha yenzwa izinyanga noma iminyaka, ngisho noma inikwe amandla ibhethri lemali.
- Izindleko eziphansi: Idizayinelwe ukuthi isebenze kuma-microcontrollers ashibhile angu-32-bit noma ama-DSP. Lezi zilawuli ezincane ngokuvamile zingamasenti ambalwa ngasinye, futhi ingqikithi yesistimu eshumekiwe eyakhiwe ngazo ingaphansi kwama-$50. Lena inketho engabizi kakhulu yokusebenzisa izinhlelo zokufunda ngomshini ngezinga elikhulu, futhi inenzuzo ikakhulukazi ezinhlelweni zokusebenza ze-IoT lapho ukufundwa komshini kufanele kusetshenziswe.
- Ukubambezeleka Okuphansi: Izinhlelo zokusebenza zinokubambezeleka okuphansi njengoba zingadingi ukuthutha noma ukushintshanisa idatha ngenethiwekhi. Yonke idatha yenzwa irekhodwa endaweni, futhi iziphetho zithathwa kusetshenziswa imodeli esivele iqeqeshiwe. Imiphumela yokucatshangwayo ingase ithunyelwe kuseva noma ifu ukuze kuganwe noma kucutshungulwe okwengeziwe, nakuba lokhu kungabalulekile ukuze idivayisi isebenze. Lokhu kunciphisa ukubambezeleka kwenethiwekhi futhi kuqeda isidingo sokuthi imisebenzi yokufunda yomshini yenziwe emafini noma kuseva.
- Inqubomgomo: Kuyinkinga enkulu ku-inthanethi kanye ne-inthanethi yezinto. Umsebenzi wokufunda ngomshini kuzinhlelo zokusebenza ze-TinyML wenziwa endaweni, ngaphandle kokugcina noma ukuthumela idatha yenzwa/yomsebenzisi kuseva/ifu. Njengomphumela, noma zixhunywe kunethiwekhi, lezi zinhlelo zokusebenza ziphephile ukusetshenziswa futhi azifaki ubungozi bobumfihlo.
Izicelo
- Ezolimo - Nini abalimi bathatha isithombe sesitshalo, uhlelo lwe-TensorFlow Lite luthola izifo kuso. Isebenza kunoma iyiphi idivayisi futhi ayidingi uxhumano lwe-inthanethi. Inqubo ivikela izintshisekelo zezolimo futhi iyisidingo esibalulekile kubalimi basemaphandleni.
- Ukugcinwa Kwemishini - I-TinyML, uma isetshenziswa kumadivayisi anamandla aphansi, ingahlala ikhomba amaphutha emshinini. Kubandakanya ukulungiswa okusekelwe ekubikezelweni. I-Ping Services, isiqalo sase-Australia, yethule igajethi ye-IoT eqapha izinjini zomoya ngokuzinamathisela ngaphandle kwe-turbine. Yazisa iziphathimandla noma nini lapho ithola noma iyiphi inkinga engaba khona noma ukungasebenzi kahle.
- Izibhedlela - The I-Solar Scare iphrojekthi. Omiyane basebenzisa i-TinyML ukuze banqande ukusabalala kwezifo ezinjengodenga nomalaleveva. Inikwa amandla amandla elanga futhi ibona izimo zokuzala komiyane ngaphambi kokubonisa amanzi ukuvimbela ukuzala komiyane.
- Ukubhekwa Kwethrafikhi - Ngu sisebenzisa i-TinyML kuzinzwa eziqoqa idatha yethrafikhi yesikhathi sangempela, singazisebenzisela ukuqondisa kangcono ithrafikhi futhi sinciphise izikhathi zokuphendula ezimotweni eziphuthumayo. I-Swim.AI, isibonelo, isebenzisa lobu buchwepheshe ekusakazeni idatha ukuze kwandiswe ukuphepha kwabagibeli kuyilapho inciphisa ukuminyana nokukhishwa kwekhabhoni ngokusebenzisa umzila ohlakaniphile.
- Law: I-TinyML ingasetshenziswa kwezomthetho ukuhlonza izenzo ezingekho emthethweni njengokuvukela umbuso nokweba kusetshenziswa ukufunda ngomshini nokubonwa kokuthinta. Uhlelo olufanayo lungasetshenziswa futhi ukuvikela ama-ATM asebhange. Ngokubuka ukuziphatha komsebenzisi, imodeli ye-TinyML ingase ibikezele ukuthi umsebenzisi ungumthengi wangempela oqedela umsebenzi noma isigebengu esizama ukugebenga noma ukucekela phansi i-ATM.
Ungaqala kanjani nge-TinyML?
Ukuze uqalise nge-TinyML ku-TensorFlow Lite, uzodinga ibhodi le-microcontroller elihambisanayo. I-TensorFlow Lite yama-Microcontrollers isekela ama-microcontroller abhalwe ngezansi.
- Itheminali ye-Wio: ATSAMD51
- I-Himax WE-I Plus EVB Endpoint AI Development Board
- Idatha ye-STM32F746
- Adafruit EdgeBadge
- I-Synopsys DesignWare ARC EM Software Development Platform
- I-Sony Spresense
- I-Arduino Nano 33 BLE Sense
- I-SparkFun Edge
- I-Adafruit TensorFlow Lite ye-Microcontrollers Kit
- I-Adafruit Circuit Playground Bluefruit
- I-Espressif ESP32-DevKitC
- I-Espressif ESP-EYE
Lawa ama-microcontroller angu-32-bit anenkumbulo eyanele ye-flash, i-RAM, nefrikhwensi yewashi ukuze asebenzise imodeli yokufunda yomshini. Amabhodi futhi anenombolo yezinzwa ezingaphakathi ezikwazi ukusebenzisa noma yiluphi uhlelo olushumekiwe nokusebenzisa amamodeli okufunda omshini kuhlelo lokusebenza oluqondisiwe. Kuya yakha imodeli yokufunda yomshini, uzodinga ikhompuyutha ephathekayo noma ikhompuyutha ngaphezu kwengxenyekazi yehadiwe.
Inkundla ngayinye yezingxenyekazi zekhompuyutha inamathuluzi ayo okuhlela okwakha, ukuqeqesha, namamodeli okufunda omshini wokuthunga, asebenzisa iphakheji ye-TensorFlow Lite ye-Microcontrollers. I-TensorFlow Lite imahhala ukuthi isetshenziswe futhi ilungiswe ngoba injalo umthombo ovulekile.
Ukuze uqalise nge-TinyML ne-TensorFlow Lite, okudingayo nje enye yezinkundla zehadiwe ezishumekiwe ezishiwo ngenhla, ikhompuyutha/ikhompyutha ephathekayo, intambo ye-USB, isiguquli se-USB-to-Serial - kanye nesifiso sokuzijwayeza ukufunda ngomshini ngamasistimu ashumekiwe. .
Izinselele
Noma inqubekelaphambili ye-TinyML iveze imiphumela eminingi emihle, imboni yokufunda ngomshini isabhekene nezingqinamba ezinkulu.
- Ukuhlukahluka kwesoftware - Ukubhala ikhodi ngesandla, ukukhiqizwa kwekhodi, nabahumushi be-ML zonke izinketho zokusebenzisa amamodeli kumadivayisi we-TinyML, futhi ngayinye ithatha inani elihlukile lesikhathi nomzamo. Ukusebenza okuhlukene kungavela ngenxa yalokhu.
- Ukuhlukahluka kwe-Hardware - Kukhona kukhona izinketho eziningana zehadiwe ezitholakalayo. Amapulatifomu e-TinyML angaba yinoma yini kusukela kuma-microcontrollers enhloso evamile kuya kuma-Cutting-Edge Neural processors. Lokhu kubangela izinkinga ngokusetshenziswa kwamamodeli kuzo zonke izinhlobo zezakhiwo.
- Ukuxazulula izinkinga / ukulungisa iphutha - Nini imodeli ye-ML ayisebenzi kahle efwini, kulula ukubuka idatha bese uthola ukuthi konakalaphi. Uma imodeli isatshalaliswa ezinkulungwaneni zamadivayisi we-TinyML, ngaphandle kokusakaza kwedatha okubuyela emafini, ukulungisa iphutha kuba nzima futhi kungase kudinge indlela ehlukile.
- Izithiyo zememori - Yendabuko amapulatifomu, afana nama-smartphone namakhompyutha aphathekayo, adinga amagigabhayithi e-RAM, kanti amadivayisi we-TinyML zisebenzisa amakhilobhayithi noma amamegabhayithi. Ngenxa yalokho, usayizi wemodeli engase isetshenziswe unomkhawulo.
- Ukuqeqeshwa kwemodeli - Nakuba kunezinzuzo ezimbalwa zokusebenzisa amamodeli e-ML kumadivayisi we-TinyML, inqwaba yamamodeli e-ML isaqeqeshwa emafini ukuze iphindaphinde futhi ithuthukise ngokuqhubekayo ukunemba kwemodeli.
Ikusasa
I-TinyML, enezinyathelo zayo ezincane, ukusetshenziswa kwebhethri okuphansi, kanye nokuntuleka noma ukuncika okulinganiselwe ekuxhumekeni kwe-inthanethi, inamandla amakhulu esikhathini esizayo, njengoba iningi lemishini emincane ukuhlakanipha okungekhona okwangempela izosetshenziswa kumadivayisi asemaphethelweni noma kumagajethi ashumekiwe azimele.
Izokwenza izinhlelo zokusebenza ze-IoT zibe yimfihlo futhi zivikeleke ngokuzisebenzisa. Nokho I-TensorFlow I-Lite okwamanje iyona yodwa uhlaka lokufunda lomshini lwama-microcontrollers nama-microcomputer, ezinye izinhlaka eziqhathanisekayo njengenzwa kanye ne-CMSIS-NN ye-ARM ziyasebenza.
Nakuba i-TensorFlow Lite iyiphrojekthi yomthombo ovulekile eqhubekayo eqale kahle ngeThimba Le-Google, isadinga ukusekelwa komphakathi ukuze ingene kwejwayelekile.
Isiphetho
I-TinyML iyindlela entsha ehlanganisa amasistimu ashumekiwe nokufunda komshini. Njengoba i-AI encane iphakama ezindaweni eziningi eziqondile nasezizindeni, ubuchwepheshe bungavela njengenkundla engezansi evelele ekufundeni komshini kanye nobuhlakani bokwenziwa.
Ihlinzeka ngesixazululo ezinseleleni eziningi umkhakha we-IoT nochwepheshe abasebenzisa ukufunda ngomshini emikhakheni eminingi eqondene nesizinda esibhekene nayo manje.
Umqondo wokusebenzisa ukufunda komshini ku amadivaysi asemaphethelweni anekhompuyutha encane i-footprint kanye nokusetshenziswa kwamandla kunamandla okuguqula kakhulu indlela amasistimu ashumekiwe namarobhothi akhiwa ngayo.
shiya impendulo