Ubuhanga bwo gukora imibonano mpuzabitsina (AI) bwabaga kuri seriveri nini na GPU zo mu bicu. Ubu burimo kugabanuka no kunyerera iruhande rw'ibikoresho byo gukora imibonano mpuzabitsina (sensors). (AI) kuri sisitemu zishyirwamo (embedded systems) si ikintu kidasanzwe - bumaze kuririmba muri frigo, drones, imyenda… ndetse n'ibikoresho bisa nkaho bitameze neza na gato.
Dore impamvu iyi mpinduka ari ingenzi, icyayigora, n'amahitamo akwiriye umwanya wawe.
Ingingo ushobora gukunda gusoma nyuma y'iyi:
🔗 Ibikoresho byiza by’imiyoborere ya AI byemeza ko sisitemu za AI zikurikiza amahame mbwirizamuco kandi zisobanutse neza
Inyigisho ku bikoresho bifasha mu kubungabunga uburiganya bushingiye ku mahame mbwirizamuco, kubahiriza amategeko, no ku mucyo.
🔗 Ububiko bw'ibintu bya AI: amahitamo, amahitamo, amahitamo
Kugereranya uburyo bwo kubika ibintu bwagenewe imirimo ya AI.
🔗 Ibisabwa mu kubika amakuru kuri AI: icyo ukeneye kumenya by'ukuri
Ibintu by'ingenzi ugomba kuzirikana mu gihe uteganya kubika amakuru ya AI.
AI kuri sisitemu zishyirwamo ibintu byinshi🌱
Ibikoresho byashyizwemo ni bito, akenshi bikoresha batiri, kandi bifite ubushobozi buke. Nyamara ubuhanga bwo gukora imibonano mpuzabitsina (AI) butuma umuntu atsinda ibintu bikomeye:
-
Ibyemezo by'igihe nyacyo nta ngendo zo kugaruka mu bicu.
-
Ubuzima bwite bushingiye ku gishushanyo - amakuru fatizo ashobora kuguma kuri igikoresho.
-
Gabanya igihe cyo gutinda iyo milisegonda ari ngombwa.
-
Gusuzuma ingufu binyuze mu guhitamo neza ibikoresho by'ikoranabuhanga.
Izi si inyungu zo gukoresha intoki gusa: gusunika mudasobwa ku nkengero bigabanya kwishingikiriza kuri interineti kandi bikongera ubuzima bwite ku bantu benshi bakoresha [1].
Amayeri si imbaraga z'ubugizi bwa nabi - ni ukuba umuhanga ufite amikoro make. Tekereza kwiruka marato ufite igikapu… maze injeniyeri zigakomeza gukuraho amatafari.
Imbonerahamwe yihuse yo kugereranya ubuhanga bwa AI kuri sisitemu zishyirwamo 📝
| Igikoresho / Umushinga | Abakiriya beza | Igiciro (hafi) | Impamvu bikora (inyandiko zidasanzwe) |
|---|---|---|---|
| TensorFlow Lite | Abakora imishinga, abakunda ibintu bitandukanye | Ubuntu | Ifite ubushobozi bwo gutwara abantu, igendanwa, ifite uburyohe bwiza bwa MCU → itumanaho rigendanwa |
| Impulse y'inkombe | Abatangizi n'ibigo bishya | Icyiciro cya Freemium | Imikorere yo gukurura no kurekura - nka "AI LEGO" |
| Urubuga rwa Nvidia Jetson | Injeniyeri zikeneye ingufu | $$$ (ntibihendutse) | GPU + accelerators zo kureba/gukora akazi gakomeye |
| TinyML (binyuze kuri Arduino) | Abarimu, abakora ibishushanyo mbonera | Igiciro gito | Yegereka; iyobowe n'abaturage ❤️ |
| Moteri ya Qualcomm AI | Abakora ibikoresho bya elegitoroniki, abakora telefoni zigendanwa | Biratandukanye | NPU yihutishijwe kuri Snapdragon - yihuta cyane |
| ExecuTorch (PyTorch) | Abahanga mu by'ikoranabuhanga rya telefoni zigendanwa n'iziciriritse | Ubuntu | Igihe cyo gukoresha PyTorch kuri telefoni/ibikoresho byambarwa/byashyizwemo [5] |
(Yego, ntaho bihuriye. Uko ni ko bimeze n'ukuri.)
Impamvu ubuhanga bwo gukora imibonano mpuzabitsina ku bikoresho byashyizwemo ari ingenzi ku nganda 🏭
Si ugutangaza gusa: ku nganda, moderi ntoya zigira inenge; mu buhinzi, utuzu duto dupima ubutaka mu murima; mu modoka, ibintu by'umutekano ntibishobora "guhamagara" mbere yo gufunga feri. Iyo gutinda no kwiherera bidashoboka , kwimura ikoranabuhanga ryo kubara ku nkengero ni uburyo bwo gufata ingamba [1].
TinyML: Intwari Ituje ya Embedded AI 🐜
TinyML ikoresha moderi kuri microcontrollers ifite kilobytes kuri megabytes nke za RAM - ariko iracyakora akazi ko gushakisha amagambo y'ingenzi, kumenya ibimenyetso, kumenya ibitagenda neza, n'ibindi. Ni nko kureba imbeba iterura itara. Birashimishije cyane.
Icyitegererezo cyihuse cyo mu mutwe:
-
Ibirenge by'amakuru : ibyuma bito bikoresha sensor yo kureba amashusho.
-
Modeli : CNNs/RNNs nto, ML ya kera, cyangwa imbuga za sparsified/quantized.
-
Ingengo y'imari : miliwatts, ntabwo ari watts; KB–MB, ntabwo ari GB.
Amahitamo y'ibikoresho: Ikiguzi ugereranyije n'imikorere ⚔️
Guhitamo ibikoresho niho imishinga myinshi ihindagurika:
-
Icyiciro cya Raspberry Pi : CPU nziza, ikoreshwa muri rusange; ikomeye ku bikoresho by'igerageza.
-
NVIDIA Jetson : modules za AI zo mu bwoko bwa "edge-byubakiwe" (urugero, Orin) zitanga TOPS nyinshi ku magana ku buryo bwo kureba neza cyangwa gukoresha modeli nyinshi - nziza, ariko ihenze cyane kandi iremereye [4].
-
Google Coral (Edge TPU) : umuvuduko wa ASIC utanga ~ 4 TOPS kuri 2W (~ 2 TOPS/W) kuri moderi zapimwe - perf/W nziza cyane iyo moderi yawe ihuye n'ibipimo [3].
-
SoCs za Smartphone (Snapdragon) : zoherezwa hamwe na NPU na SDK kugira ngo zikore neza moderi kuri terefone.
Itegeko rigenga: kuringaniza ikiguzi, ubushyuhe, n'imibare. "Byiza bihagije, ahantu hose" akenshi biruta "ubuhanga buhanitse, ntaho bihuriye."
Imbogamizi zisanzwe muri AI kuri sisitemu zishyirwamo 🤯
Injeniyeri zihora zihanganye na:
-
Ubwibutso buke : ibikoresho bito ntibishobora kwakira moderi nini.
-
Ingengo y'imari ya batiri : buri milimita ni ingenzi.
-
Kunoza icyitegererezo:
-
Ingano → uburemere/ibikorwa bito, byihuse bya int8/float16.
-
Gukata → gukuraho ibintu biremereye bito cyane kugira ngo ubone umusaruro muke.
-
Gushyira hamwe/gusangiza ibiro → gukanda cyane.
Izi ni tekiniki zisanzwe zo gukoresha neza ibikoresho [2].
-
-
Kongera ubushobozi : icyitegererezo cya Arduino mu ishuri ≠ sisitemu yo gukora imodoka ifite umutekano, umutekano, n'imbogamizi ku buzima.
Urimo gukosora amakosa? Tekereza usoma igitabo unyuze mu mwobo w'urufunguzo… ufite ingofero.
Porogaramu zifatika uzabona izindi vuba 🚀
-
Imyenda igezweho ikora ubushakashatsi ku buzima bw'ibikoresho.
-
Kamera za IoT zigaragaza ibikorwa bidakoresheje amashusho y’umwimerere.
-
Abafasha b'ijwi batari kuri interineti kugira ngo bagenzure intoki badakoresheje intoki - nta kwishingikiriza ku bicu.
-
Indege zitagira abapilote zikoreshwa mu kugenzura, gutanga no gukora neza.
Muri make: Ubuhanga bwo gukora imibonano mpuzabitsina burimo kwegera cyane - mu maboko yacu, mu gikoni cyacu, no mu bikorwa remezo byacu.
Uburyo Abakora Inganda Bashobora Gutangira 🛠️
-
Tangira na TensorFlow Lite kugira ngo ukoreshe ibikoresho byinshi kandi ukoreshe MCU → mobile coverage; shyiramo quantization/curuning hakiri kare [2].
-
Suzuma ExecuTorch niba utuye mu gace ka PyTorch kandi ukeneye igihe cyo gukoresha ikoranabuhanga kuri telefoni zigendanwa n’izishyizwemo [5].
-
Gerageza ibikoresho bya Arduino + TinyML kugira ngo ubone ibishushanyo byihuse kandi bishimishije.
-
Ukunda imiyoboro igaragara? Edge Impulse igabanya inzitizi ikoresheje gufata amakuru, kuyahugura no kuyashyira mu bikorwa.
-
Fata ibikoresho nk'icyitegererezo cy'umuturage wo mu rwego rwa mbere - prototype kuri CPU, hanyuma wemeze kuri accelerator yawe (Edge TPU, Jetson, NPU) kugira ngo wemeze latency, thermal, na accuracy deltas.
Mini-vignette: Itsinda ryohereza icyuma gipima vibration-anomaly kuri sensor ya selile y'ibiceri. Modeli ya float32 ibura ingengo y'imari y'amashanyarazi; int8 quantization igabanya ingufu kuri buri gipimo, igabanya kwibuka, kandi MCU irangiza akazi - nta muyoboro ukenewe [2,3].
Impinduramatwara ituje ya AI kuri sisitemu zishyirwamo 🌍
Imashini nto kandi zihendutse zirimo kwiga kumva → gutekereza → gukora - aho ziri. Ubuzima bwa bateri buzahora budukurikirana, ariko inzira irasobanutse: moderi zikomeye, compilers nziza, accelerators nziza. Ingaruka ni izihe? Ikoranabuhanga rirushaho kuba ryiza kandi risubiza ibibazo kuko ritahujwe gusa - ni ukwita ku byo rikora.
Amareferensi
[1] ETSI (Multi-access Edge Computing) - Inyungu zo gutinda/kwishingira ubuzima bwite n'imiterere y'inganda.
ETSI MEC: Incamake nshya y'inyandiko yera
[2] Google TensorFlow Model Optimization Toolkit - Gupima, gukata, gukusanya amakuru kugira ngo igikoresho gikore neza.
Ubuyobozi bwo Gutunganya TensorFlow Model
[3] Google Coral Edge TPU - Ibipimo bya Perf/W byo kwihutisha imipaka.
Ibipimo bya Edge TPU
[4] NVIDIA Jetson Orin (yemewe) - Edge AI modules n'ibipimo by'imikorere.
Incamake ya Jetson Orin Modules
[5] PyTorch ExecuTorch (Inyandiko zemewe) - Igihe cyo gukoresha PyTorch kuri telefoni na interineti.
Incamake ya ExecuTorch