Isiqulatho[Fihla][Bonisa]
- Yintoni ukuhlelwa kwemifanekiso?
- Ukuhlelwa kwemifanekiso kusebenza njani?
Ukuhlelwa kwemifanekiso kusetyenziswa iTensorflow & Keras kwipython+-
- 1. IiMfuno zokuFakela
- 2. Ukungenisa abaxhomekeke
- 3. Ukuqalisa iiparamitha
- 4. Ilayisha isethi yedatha
- 5. Ukudala imodeli
- 6. Ukuqeqesha imodeli
- Ukuvavanywa kwemodeli
- 7. Ukungenisa izinto eziluncedo kuvavanyo
- 8. Ukwenza uluhlu lwepython
- 9. Ukulayisha idatha yovavanyo kunye nemodeli
- 10. UVavanyo nokuQikelela
- 11. Iziphumo
- isiphelo
Kuyakhuthaza ukwazi ukuba sikwazile ukufaka iirobhothi ngobuchule bethu bendalo bokufunda ngomzekelo kunye nokuqonda indawo ezingqongileyo. Owona mceli mngeni ungundoqo ngowokufundisa iikhomputha ukuba “zibone” njengokuba abantu beya kufuna ixesha elininzi kunye nomzamo.
Nangona kunjalo, xa sicinga ngexabiso elisebenzayo olu buchule bubonelela ngoku kwimibutho nakumashishini, umgudu ufanelekile. Kweli nqaku, uya kufunda malunga nokuhlelwa kwemifanekiso, indlela esebenza ngayo, kunye nokuphunyezwa kwayo okusebenzayo. Masiqale.
Yintoni ukuhlelwa kwemifanekiso?
Umsebenzi wokondla umfanekiso ube a inethiwekhi yomnatha kwaye ukuba ikhuphe uhlobo oluthile lweleyibhile yalo mfanekiso yaziwa njengoluqwalaselo lomfanekiso. Ileyibhile yemveliso yenethiwekhi iya kuhambelana nodidi oluchazwe kwangaphambili.
Kusenokubakho iiklasi ezininzi ezabelwe umfanekiso, okanye enye nje. Xa kukho iklasi enye kuphela, igama elithi "ukuqaphela" lisetyenziswa rhoqo, ngelixa kukho iiklasi ezininzi, igama elithi "uhlelo" lisetyenziswa rhoqo.
Ukufunyanwa kwento yiseti yokuhlelwa kwemifanekiso apho iimeko ezithile zezinto ziye zafunyanwa njengezizezodidi olunikiweyo njengezilwanyana, izithuthi, okanye abantu.
Ukuhlelwa kwemifanekiso kusebenza njani?
Umfanekiso okwimo yeepixels uhlalutywa yikhompyuter. Ifezekisa oku ngokuphatha umfanekiso njengengqokelela yeematriki, ubungakanani bayo buchongiwe sisisombululo somfanekiso. Ukuthetha nje, ukuhlelwa kwemifanekiso kuphononongo lwedatha yamanani kusetyenziswa i-algorithms ngokwembono yekhompyuter.
Ukuhlelwa kwemifanekiso kufezekiswa ekusetyenzweni kwemifanekiso yedijithali ngokuhlanganisa iipixels zibe ngamaqela amiselweyo, okanye “iiklasi.” Ii-algorithms zahlula umfanekiso ngokulandelelana kweempawu eziphawulekayo, ezinciphisa umthwalo kumhleli wokugqibela.
Ezi mpawu zazisa umahluli malunga nentsingiselo yomfanekiso kunye nokuhlelwa okunokwenzeka. Ngenxa yokuba ezinye iinkqubo zokuhlelwa komfanekiso zixhomekeke kuyo, indlela yokutsalwa kweempawu sesona sigaba sibaluleke kakhulu.
The idatha enikezelweyo kwi-algorithm ikwabalulekile kuhlelo lwemifanekiso, ngakumbi ukuhlelwa okugadiweyo. Xa kuthelekiswa neseti yedatha eyoyikekayo kunye nokungalingani kwedatha okusekwe kwiklasi kunye nomfanekiso ophantsi kunye nomgangatho wenkcazo, isethi yedatha yokuhlelwa kakuhle isebenza ngokuncomekayo.
Ukuhlelwa kwemifanekiso kusetyenziswa iTensorflow & Keras kwipython
Siza kusebenzisa i I-CIFAR-10 idataset (ebandakanya iinqwelomoya, iinqwelomoya, iintaka, kunye nezinye izinto ezisi-7).
1. IiMfuno zokuFakela
Ikhowudi engezantsi iya kufaka zonke izinto ezifunekayo kuqala.
2. Ukungenisa abaxhomekeke
Yenza ifayile ye-train.py kwiPython. Ikhowudi engezantsi iya kungenisa iTensorflow kunye nokuxhomekeka kweKeras.
3. Ukuqalisa iiparamitha
I-CIFAR-10 ibandakanya iindidi zemifanekiso ezili-10, kungoko iiklasi zenani zibhekisa nje kwinani leendidi zokuhlela.
4. Ilayisha isethi yedatha
Umsebenzi usebenzisa imodyuli yeTensorflow Datasets ukulayisha iseti yedatha, kwaye siseta ngolwazi ukuya kwiNyaniso ukufumana ulwazi malunga nayo. Ungayiprinta ukuze ubone ukuba yeyiphi imimandla kunye nexabiso layo, kwaye siya kusebenzisa ulwazi ukufumana kwakhona inani leesampuli kuqeqesho kunye neeseti zovavanyo.
5. Ukudala imodeli
Ngoku siza kwakha iileya ezintathu, nganye iqulathe iiConvNets ezimbini ezinodibaniso olukhulu kunye nomsebenzi wokuvula i-ReLU, ilandelwa yinkqubo yeyunithi ye-1024 eqhagamshelwe ngokupheleleyo. Xa kuthelekiswa ne-ResNet50 okanye i-Xception, eziyimodeli yangoku, le inokuba yimodeli encinci.
6. Ukuqeqesha imodeli
Ndisebenzise i-Tensorboard ukulinganisa ukuchaneka kunye nelahleko kwixesha ngalinye kwaye ndisinike umboniso othandekayo emva kokungenisa idatha kunye nokuvelisa imodeli. Qhuba le khowudi ilandelayo; kuxhomekeke kwi-CPU/GPU yakho, uqeqesho luya kuthatha imizuzu emininzi.
Ukusebenzisa i-tensorboard, chwetheza nje lo myalelo ulandelayo kwi-terminal okanye umyalelo womyalelo kulawulo lwangoku:
Uya kubona ukuba ilahleko yokuqinisekisa iyancipha kwaye ukuchaneka kunyuka ukuya malunga ne-81%. Yinto entle leyo!
Ukuvavanywa kwemodeli
Xa uqeqesho lugqityiwe, imodeli yokugqibela kunye neentsimbi zigcinwa kwifolda yeziphumo, okusivumela ukuba siqeqeshe kube kanye kwaye senze uqikelelo nanini na xa sikhetha. Landela ikhowudi kwifayile entsha yepython ebizwa ngokuba yi-test.py.
7. Ukungenisa izinto eziluncedo kuvavanyo
8. Ukwenza uluhlu lwepython
Yenza isichazi-magama sePython esiguqulela ixabiso ngalinye elipheleleyo kwileyibhile efanelekileyo yesethi yedatha:
9. Ukulayisha idatha yovavanyo kunye nemodeli
Ikhowudi elandelayo iya kulayisha idatha yovavanyo kunye nemodeli.
10. UVavanyo nokuQikelela
Ikhowudi elandelayo iya kuvavanya kwaye yenza uqikelelo kwimifanekiso yesele.
11. Iziphumo
Imodeli yaxela kwangaphambili isele ngokuchaneka okungama-80.62%.
isiphelo
Kulungile, sigqibile ngesi sifundo. Ngelixa i-80.62% ingayilungelanga i-CNN encinci, ndikucebisa ngamandla ukuba uguqule imodeli okanye ujonge i-ResNet50, i-Xception, okanye ezinye iimodeli ezijonge phambili kwiziphumo ezingcono.
Ngoku ekubeni wakhe inethiwekhi yakho yokuqala yokuqatshelwa kwemifanekiso eKeras, kuya kufuneka ulinge imodeli ukufumanisa ukuba iiparamitha ezahlukeneyo ziyichaphazela njani ukusebenza kwayo.
Shiya iMpendulo