InoteveraGaidhi rinotevera
Bias Mitigation Techniques: Pre-, In- uye Post-Processing
Tekinoroji
Nhungamiro yehunyanzvi
A neural processing unit, kana NPU, inyanzvi hardware yekumhanyisa neural-network computation.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Mumafoni nemalaptops inogona kubatsira kumhanya inotsigirwa inference workloads zvinobudirira padivi peCPU neGPU. Kubatsira kwayo kunoenderana nemuenzaniso, software, ndangariro, uye kuyerwa kwekuita basa, kwete kungoshambadzirwa kwekuita basa.
An NPU ndeye computing accelerator, kwete yakadzidziswa modhi kana yakasarudzika njere mukati memudziyo. Neural network inoita akadzokororwa manhamba mashandiro, kusanganisira matrix kuwanza uye convolution. Specialized hardware inogona kuronga mashandiro aya uye kufambisa data nemazvo kune inotsigirwa mitoro yebasa. Arm inotsanangura maNPU seanosimudzira neural-network inference; Zvinyorwa zveIntel's NPU zvakare zvinotsanangura zvakatsaurirwa mabhuroki emakomputa uye compiler inoronga basa uye data kufamba. Iyo CPU, GPU, uye NPU inogona kuva nemabasa anowirirana. A CPU inomhanya zvakajairika application logic, nepo accelerator inobata basa iro software yayo stack uye hardware rutsigiro. Kuve neNPU hakumanikidze yega yega AI chimiro kuti imhanye ipapo. Chimwe chishandiso chinogona kushandisa CPU, GPU, sevhisi iri kure, kana musanganiswa. Tarisa maitiro ekushandisa chaiwo usati wapa mhedzisiro kuNPU. Kuenderana kunonyanya kujeka kupfuura yekuwedzera faira remuenzaniso. Maoperator, maumbirwo, manhamba mafomati, ndangariro zvinodiwa, madhiraivha, uye runtime rutsigiro zvinokanganisa kutumirwa. Nguva yekumhanya inogona kupatsanura modhi girafu pakati peanotsigirwa execution backends; ONNX Runtime magwaro ekugona-kwakavakirwa kugoverwa kwemanodhi uye subgraphs. Basa risingatsigirwe rinogona kushandisa imwe yakagadziridzwa backend kana kudzivirira gadziriro kuti isashande. Ongorora matanda uye profiling pane kufungidzira kuburitsa kwakabudirira zvinoreva kukwidziridza kuzere. Enzanisa basa rinotarisirwa pahutano hunodiwa. Kune vhidhiyo mhedzisiro, ongorora furemu latency, simba, hunhu hwakasimba, uye zvinoonekwa zvinobuda. Kuti unyore, sanganisira kuzivikanwa mhando uye iyo yakazara yekuteerera-yekugadzirisa nzira. Kana shanduko kana quantization ichichinja hunhu hwenhamba, ongorora mhedzisiro modhi. Kuita munharaunda kunogona kuderedza kumwe kutamiswa kwedata, asi maficha akatenderedza anogona kuramba achiisa matanda kana kuwiriranisa mhinduro. Iyo chip nzvimbo haigadzirise kuvanzika maitiro echishandiso chose.
Yekufungidzira vhidhiyo-kufona application inoshandisa inotsigirwa kumashure-segmentation modhi paNPU nepo CPU ichibata application. Chikwata chinoyera simba uye furemu latency panguva yekufona chaiko.
Kamera application inoshandisa yakashandurwa mufananidzo modhi, asi oparesheni imwe chete haina kutsigirwa neyakasarudzwa nzira yekumhanyisa. Vagadziri vanoongorora nguva yekuiswa panzvimbo pekufungidzira kuti modhi yese yakamhanya paNPU.
Mutengi anotarisa kana iyo yekunyorera application yavanoda inotsigira laptop's NPU. Kuvapo kwechip chete hakuratidze kuti iyi application ichaishandisa.
Chikwata chinofananidza modhi isati yasvika uye mushure mekushandura yakaderera-yechokwadi, kuyedza ese ari maviri emhando yebasa uye kushandiswa kwechishandiso pane chakagadzirirwa mudziyo.
NPU Hardware nerutsigiro rwesoftware ichaenderera mberi ichichinja kuzvizvarwa zvemidziyo. Mamwe mabasa anotsigirwa kana kuunganidzwa kuri nani kunogona kuita kuti mamwe mabasa ashande, asi kuenderana kwekushandisa kunofanirwa kutariswa zvakare mushure memuenzaniso kana nguva yekumhanyisa. Chengetedza diki bvunzo seti uye rekodhi yekuitwa kuiswa, mhando, latency, uye simba kune izvo zvakakosha. Kukwidziridzwa kunobatsira kunovandudza iro chairo basa mukati mezvipingaidzo zvaro. Izvo hazvifanirwe kutongwa chete nenhamba yakakura yepamusoro kana kuti chigadzirwa chakanyorwa chine izwi rekuti AI.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
A neural processing unit, kana NPU, inyanzvi hardware yekumhanyisa neural-network computation. Mumafoni nemalaptops inogona kubatsira kumhanya inotsigirwa inference workloads zvinobudirira padivi peCPU neGPU. Kubatsira kwayo kunoenderana nemuenzaniso, software, ndangariro, uye kuyerwa kwekuita basa, kwete kungoshambadzirwa kwekuita basa.
Nhungamiro inotsanangura NPU seyekumhanyisa, yakaparadzana nemuenzaniso uye kushandisa kuishandisa.
Kuvapo kweNPU hakuratidze kuti imwe application inoishandisa.
Gwaro rinotsanangura kugona-kwakavakirwa kugoverwa kwezvikamu zvegraph kune zvinotsigirwa kumashure.
TOPS chikamu chekushanda-chiyero. Kufanotaura kunowanzo sanganisira mabasa mazhinji pamwe nemamwe mabasa.
Nhungamiro inoti kutendeuka kana quantization inogona kushandura hunhu hwenhamba, saka mhedzisiro modhi inofanirwa kuongororwa.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Bias Mitigation Techniques: Pre-, In- uye Post-Processing
Tekinoroji