Mazhinji ekudzidza muchina uye akadzama ekudzidza modhi anovimba zvakanyanya nehuwandu hwe data uye akasiyana kuti ashande nemazvo. Huwandu uye kusiyana-siyana kwedata rakapihwa panguva yekudzidziswa zvine chekuita kwakakura pakufanotaura chokwadi chemhando idzi.
Mienzaniso yekudzidza yakadzama iyo yakadzidziswa kuita zvinobudirira pamabasa akaoma kazhinji inosanganisira yakavanzika neuroni. Huwandu hwema paramita anodzidziswa hunowedzera zvichienderana nenhamba yakavanzika neuroni.
Huwandu hwe data hunodiwa hunoenderana nenhamba yemuenzaniso inodzidziswa paramita. Imwe nzira yekubata nekuomerwa kwe data shoma ndeyekushandisa akasiyana shanduko kune yazvino data kugadzira data nyowani.
Iyo nzira yekugadzira data nyowani kubva kune iripo data inonzi 'Data Augmentation.' Kuwedzeredzwa kwedata kunogona kushandiswa kuzadzisa zvese zvinodiwa: huwandu hwe data uye akasiyana eiyo data yekudzidziswa inodiwa kuti ugadzire chokwadi. Kudzidza kwemichina kana modhi yekudzidza yakadzama.
Mune ino positi, isu tinotarisa zvakanyanya kuwedzera data, mhando dzayo, nei ichikosha, uye zvimwe zvakawanda.
Saka, chii chinonzi Data Augmentation?
Data Augmentation ndiyo nzira yekugadzira itsva uye inomiririra data kubva kune iripo data. Iwe unogona kuita izvi nekubatanidza yakagadziridzwa vhezheni yedata iripo kana synthesizing data nyowani.
Iwo ma dataset anogadzirwa nenzira iyi anovandudza kudzidza kwemuchina wako kana mienzaniso yekudzidza yakadzama nekuderedza njodzi yekuwandisa. Ndiyo maitiro ekuchinja, kana "kuwedzera," dataset ine rumwe ruzivo.
Iyi yekuwedzera yekuwedzera inogona kubva pamifananidzo kuenda kune zvinyorwa, uye inosimudzira mashandiro emuchina kudzidza masisitimu.
Fungidzira kuti tinoda kuvaka modhi yekuisa mumapoka embwa uye isu tine nhamba huru yemifananidzo yemarudzi ese kunze kwemapugs. Nekuda kweizvozvo, iyo modhi yaizonetseka kurongedza pugs.
Tinogona kuwedzera mamwe (chaiwo kana emanyepo) mapikicha epug kuunganidzwa, kana isu tinokwanisa kupeta kaviri yedu yazvino mafoto epug (semuenzaniso nekudzokorora uye nekuakanganisa kuti aite akasiyana).
Chii chinoshandiswa nekuwedzera data panguva ino?
Applications for machine learning ari kukurumidza kukura uye akasiyana, kunyanya mumunda wekudzidza kwakadzama. Matambudziko anotarisana neindasitiri yehungwaru anogona kukundwa kuburikidza nemaitiro ekuwedzera data.
Kuwedzerwa kwedata kunogona kuvandudza mashandiro uye mhedzisiro yemamodhi ekudzidza muchina nekuwedzera mitsva uye yakasiyana mienzaniso kumaseti ekudzidzisa.
Kana iyo dataset yakakura uye yakakwana, modhi yekudzidza yemuchina inoita zvirinani uye yakanyatsojeka. Kumamodheru ekudzidza muchina, kuunganidza data nekuisa mazita kunogona kutora nguva uye kudhura.
Makambani anogona kuderedza mari yavo yekushanda nekushandura dheta uye kushandisa nzira dzekuwedzera data.
Kuchenesa data ndeimwe yematanho mukugadzirwa kwemuenzaniso we data, uye inokosha kune yakakwirira-yakarurama mienzaniso. Nekudaro, iyo modhi haizokwanise kutarisira zvakaringana kubva kunyika chaiyo kana kucheneswa kwedata kuchideredza kumiririrwa.
Modhi yekudzidza yemuchina inogona kusimbiswa nekushandisa nzira dzekuwedzera data, izvo zvinoburitsa misiyano ingasangana nemuenzaniso munyika chaimo.
Mhando dzeData Augmentation
Real data augmentation
Kuwedzera kwedata chaiko kunoitika kana iwe ukawedzera chaiyo, yekuwedzera data kune dataset. Izvi zvinogona kubva pamafaira ezvinyorwa ane humwe hunhu (yemifananidzo yakamaka) kuenda kumifananidzo yezvimwe zvinhu inofananidzwa nechinhu chepakutanga, kana kunyange kurekodha kwechinhu chaicho.
Semuenzaniso, nekuwedzera zvimwe zvishoma kune faira remufananidzo, muchina-wekudzidza modhi unogona kuona chinhu chacho zviri nyore.
Mamwe metadata nezvemufananidzo wega wega (semuenzaniso, zita rawo uye tsananguro) inogona kuverengerwa kuitira kuti yedu AI modhi izive zvakawanda nezve chinomiririrwa nemufananidzo wega wega usati watanga kudzidziswa pamifananidzo iyoyo.
Kana yasvika nguva yekuisa mumapoka mafoto matsva mune chimwe chezvikamu zvedu zvakafanotemerwa, senge "katsi" kana "imbwa," modhi inogona kukwanisa kuona zvinhu zviri mumufananidzo uye kuita zvirinani semhedzisiro.
Chakaita zvokugadzirwa Data Wedzera
Kunze kwekuwedzera mamwe data chaiwo, iwe unogona zvakare kupa synthetic data kana data rekugadzira rinoratidzika kunge rechokwadi.
Izvi zvinobatsira kumabasa akaoma senge neural style transfer, asi zvakare yakanakira chero dhizaini, ingave uri kushandisa maGANs (Generative Adversarial Networks), CNNs (Convolutional Neural Networks), kana mamwe akadzika neural network architecture.
Semuyenzaniso, kana tichida kunyatso patsanura pugs pasina kubuda uye kutora akati wandei mapikicha, tinogona kuwedzera mamwe mapikicha enhema muunganidzwa wemifananidzo yembwa.
Iyi fomu yekuwedzera data inonyanya kushanda mukusimudzira modhi chaiyo kana kuunganidza data kwakaoma, kudhura, kana kutora nguva. Mumamiriro ezvinhu aya, tiri kuwedzera zvinyorwa zve dataset.
Fungidzira kuti boka redu rekutanga remifananidzo 1000 yerudzi rwembwa rine chete 5 pug mifananidzo. Panzvimbo pekuwedzera mamwe mapikicha chaiwo epug kubva kumbwa chaidzo, ngatigadzirei yekunyepera nekugadzira imwe yeazvino uye nekuikanganisa zvishoma kuti irambe ichiita senge pug.
Data Augmentation Techniques
Maitiro ekuwedzera data anosanganisira kuita zvishoma zvigadziriso kune iripo data. Zvakangofanana nekudzokorora chirevo. Tinogona kupatsanura kuwedzera kwedata muzvikamu zvitatu:
Text
- Kutsiviwa kweIzwi: Iyi nzira yekuwedzera data inosanganisira kutsiva mazwi azvino nemasinonymi. Semuenzaniso, "Iyi firimu ibenzi" inogona kuve "Iyi firimu idiotic."
- Mutsara/Kukwenya Kwezwi: Iri zano rinosanganisira kushandura kutevedzana kwemitsara kana mazwi uchichengeta kuwirirana kwakazara.
- Syntax-Muti Manipulation: Unoshandura mutsara uripo kuti uve wakarurama mugirama uchishandisa mazwi mamwe chete.
- Random Deletion: Kunyangwe zano iri richigadzira kunyora kwakashata, rinoshanda. Nekuda kweizvozvo, mutsara wekuti "ini handisi kuzotenga rekodhi iyi nekuti yakakweshwa" inova "Ini handisi kuzotenga izvi nekuti zvakakweshwa." Chirevo chacho hachina kujeka, asi chinoramba chiri chekuwedzera kunonzwisisika.
- Back Translation: Iyi nzira inoshanda uye inonakidza. Tora chirevo chakanyorwa mumutauro wako, chishandurire kune mumwe mutauro, uye wochishandura zvakare uchidzokera kumutauro wako wekutanga.
Images
- Kernel Filters: Iyi nzira inorodza kana kudzima mufananidzo.
- Mufananidzo Musanganiswa: Kunyangwe zvingaite sezvisinganzwisisike, unogona kusanganisa mafoto.
- Erasing at Random: Delete chidimbu chidiki chemufananidzo wazvino.
- Shanduko dzeGeometric: Iyi nzira inosanganisira, pakati pezvimwe zvinhu, kutenderedza, kutenderedza, kuchekerera, kana kuturikira mifananidzo.
- Kutenderedza mufananidzo: Unogona kuchinjisa mufananidzo kubva kune yakachinjika kuenda kune yakatwasuka.
- Ruvara Space Shanduko: Unogona kushandura iyo RGB mavara chiteshi kana kuwedzera chero yazvino ruvara.
- Re-Scaling ndiyo maitiro ekugadzirisa chiyero chekuona. Iwe une sarudzo yekuwedzera mukati kana kunze. Paunoyera mukati, chifananidzo chinova chidiki pane saizi yekutanga. Mufananidzo unenge wakakura kudarika wepakutanga kana ukauyera kunze.
Audio
- Pitch: Iyi nzira inosanganisira kushandura inzwi rekuteerera.
- Shandura kumhanya: Shandura kumhanya kweodhiyo faira kana kurekodha.
- Ruzha rwakawanda: Unogona kuwedzera ruzha kune faira rekuteerera.
Shandisa Nyaya
Kufungidzira kwekurapa inyaya yakakurumbira yekushandisa yekuwedzera data izvozvi. Kuunganidzwa kwemifananidzo yekurapa idiki, uye kugovana data kwakaoma nekuda kwemitemo uye nezvekuvanzika.
Uyezve, seti dzedata dzakanyanya kumanikidzwa mumamiriro ezvinhu asina kujairika kusagadzikana. Makambani ekufungidzira ekurapa anoshandisa data kuwedzera kusiyanisa yavo data seti.
matambudziko
Scalability, akasiyana dhataseti, uye kukosha ndezvimwe zvezvinhu zvinoda kugadziriswa kuti ugadzire hunyanzvi hwekuwedzera data.
Panyaya ye scalability, data yakawedzerwa inofanirwa kuve yakasimudzwa kuitira kuti akawanda akasiyana mamodheru anogona kuishandisa. Iwe unozoda kuve nechokwadi chekuti izvi zvinogona kudzokororwa kuti zvishandiswe mune ramangwana mamodheru kubvira kumisikidza data yekuwedzera sisitimu iyo inoburitsa yakakura yakawanda yakakosha, yakakosha, yakakwidziridzwa data inogona kutora nguva.
Panyaya yehterogeneity, akasiyana dhataseti ane akasiyana maficha anofanirwa kutariswa uchigadzira yakawedzera data. Kugadzira data rakawedzerwa rakakodzera, zvimiro zvega yega dataset zvinofanirwa kushandiswa.
Mune mamwe mazwi, kuwedzera kwedata kuchasiyana pakati pe dataset uye kesi dzekushandisa.
Chekupedzisira, kuvimbisa kuti mabhenefiti eiyo data yakawedzera anopfuura chero njodzi, iyo yakawedzerwa data inofanirwa kuongororwa uchishandisa akakodzera metrics isati yashandiswa nemamodhi ekudzidza muchina.
Semuenzaniso, kuvapo kweruzha rwekumashure kana zvinhu zvisingaenderane mumifananidzo-based augmented data zvinogona kukanganisa kuita kwemuenzaniso.
mhedziso
Pakupedzisira, kunyangwe uri kuyedza kufanotaura kurasikirwa, kuona hutsotsi hwemari, kana kuvaka zvirinani. mufananidzo classification mamodheru, kuwedzera data inzira yakakosha yekuvaka mamwe akanyanya, akasimba modhi.
Kuburikidza nemaitiro epamusoro ekudzidzisa, nyore preprocessing uye kuwedzera data kunogona kubatsira zvikwata mukugadzira ekucheka-kumucheto modhi.
Mabhizinesi anogona kushandisa kuwedzera data kudzikisa huwandu hwenguva inopedzerwa kugadzirira dhata rekudzidziswa uye kugadzira michina yekudzidza mhando dzakanyanya uye nekukurumidza..
Nekuwedzera huwandu hwe data rakakosha mune dataset, kuwedzera data kunogona kubatsira zvakare mamodhi ekudzidza emuchina atove nedata rakawanda.
Leave a Reply