M'ndandanda wazopezekamo[Bisani][Show]
Ndizolimbikitsa kudziwa kuti takwanitsa kuyika maloboti ndi luso lathu lobadwa nalo kuti tiphunzire mwachitsanzo ndikuzindikira malo omwe ali. Vuto lalikulu ndi lophunzitsa makompyuta kuti "awone" monga momwe anthu amafunira nthawi yochulukirapo komanso khama.
Komabe, tikaganizira za phindu lomwe lusoli limapereka panopa ku mabungwe ndi mabizinesi, kuyesetsako kumakhala kopindulitsa. M'nkhaniyi, muphunzira za m'magulu azithunzi, momwe zimagwirira ntchito, komanso momwe zimakhalira. Tiyeni tiyambe.
Kodi gulu la zithunzi ndi chiyani?
Ntchito yodyetsa fano kukhala a neural network ndikupangitsa kuti chitulutse chizindikiro cha chithunzicho chimatchedwa kuzindikira zithunzi. Zolemba za netiweki zidzagwirizana ndi kalasi yofotokozedwatu.
Pakhoza kukhala makalasi ambiri operekedwa pachithunzichi, kapena limodzi lokha. Pakakhala gulu limodzi lokha, mawu oti "kuzindikira" amagwiritsidwa ntchito pafupipafupi, pomwe pakakhala magulu angapo, mawu oti "gulu" amagwiritsidwa ntchito pafupipafupi.
Kuzindikira kwa chinthu ndi kagawo kakang'ono kazithunzi komwe nthawi zina za zinthu zimazindikirika kuti ndi za gulu lomwe laperekedwa monga nyama, magalimoto, kapena anthu.
Kodi kugawa zithunzi kumagwira ntchito bwanji?
Chithunzi chokhala ngati ma pixel chimawunikidwa ndi kompyuta. Imakwaniritsa izi pochiza chithunzicho ngati chophatikiza cha matrices, kukula kwake komwe kumatsimikiziridwa ndi chithunzithunzi. Mwachidule, gulu la zithunzi ndikuwerenga ziwerengero pogwiritsa ntchito ma algorithms kuchokera pamakompyuta.
Kugawa zithunzi kumatheka pokonza zithunzi za digito poyika ma pixel kukhala magulu odziwikiratu, kapena "makalasi." Ma algorithms amagawanitsa chithunzicho motsatizana ndi mikhalidwe yodziwika bwino, zomwe zimachepetsa mtolo wa womaliza womaliza.
Makhalidwewa amadziwitsa wogawa za tanthauzo lachithunzicho komanso momwe angagawire. Chifukwa njira zina zogawira chithunzi zimadalira, njira yochotsera mawonekedwe ndiyo gawo lovuta kwambiri.
The deta yoperekedwa ku algorithm ndiyofunikiranso pakugawika kwazithunzi, makamaka gulu loyang'aniridwa. Poyerekeza ndi seti yoyipa yokhala ndi kusalinganiza kwa data kutengera kalasi ndi chithunzi chotsika ndi mtundu wa mawu, gulu lokonzekera bwino lomwe limagwira ntchito modabwitsa.
Kugawa kwazithunzi pogwiritsa ntchito Tensorflow & Keras mu python
Tidzagwiritsa ntchito CIFAR-10 dataset (yomwe imaphatikizapo ndege, ndege, mbalame, ndi zinthu zina 7).
1. Kuyika Zofunikira
Khodi ili m'munsiyi ikhazikitsa zonse zofunika.
2. Kutengera zomwe zimadalira
Pangani fayilo ya train.py ku Python. Khodi ili pansipa idzalowetsa Tensorflow ndi Keras kudalira.
3. Kuyambitsa magawo
CIFAR-10 imaphatikizanso magawo 10 a zithunzi, chifukwa chake makalasi owerengera amangotanthauza kuchuluka kwamagulu oti muwagawire.
4. Kutsegula deta
Ntchitoyi imagwiritsa ntchito gawo la Tensorflow Datasets kutsitsa deta, ndipo timayika zambiri ku True kuti tidziwe zambiri za izo. Mutha kuzisindikiza kuti muwone zomwe zili ndi makonda ake, ndipo tidzagwiritsa ntchito chidziwitsocho kuti tipeze kuchuluka kwa zitsanzo mumaphunziro ndi mayeso.
5. Kupanga chitsanzo
Tsopano tipanga zigawo zitatu, iliyonse ili ndi ma ConvNets awiri okhala ndi max-pooling ndi ReLU activation ntchito, yotsatiridwa ndi makina olumikizidwa kwathunthu a 1024-unit. Poyerekeza ndi ResNet50 kapena Xception, zomwe ndi zitsanzo zamakono, izi zitha kukhala zachitsanzo chaching'ono.
6. Kuphunzitsa chitsanzo
Ndidagwiritsa ntchito Tensorboard kuyeza kulondola ndi kutayika mu nthawi iliyonse ndikutipatsa chiwonetsero chokongola titatha kuitanitsa deta ndikupanga mtunduwo. Thamangani nambala iyi; kutengera CPU/GPU yanu, maphunziro atenga mphindi zingapo.
Kuti mugwiritse ntchito tensorboard, ingolembani lamulo lotsatirali mu terminal kapena command prompt m'ndandanda wapano:
Mudzawona kuti kutayika kovomerezeka kukucheperachepera ndipo kulondola kukukwera pafupifupi 81%. Ndizodabwitsa!
Kuyesa chitsanzo
Maphunziro akatha, chitsanzo chomaliza ndi zolemera zimasungidwa mufoda ya zotsatira, kutilola kuti tiphunzitse kamodzi ndikupanga maulosi nthawi iliyonse yomwe tasankha. Tsatirani kachidindo mu fayilo yatsopano ya python yotchedwa test.py.
7. Kutumiza zida zoyesera
8. Kupanga bukhu la python
Pangani dikishonale ya Python yomwe imamasulira chiwerengero chilichonse kukhala lebulo loyenera la dataset:
9. Kutsegula deta yoyesera & chitsanzo
Khodi yotsatirayi idzanyamula deta yoyesera ndi chitsanzo.
10. Kuunika & Kuneneratu
Khodi yotsatirayi iwunika ndikulosera pazithunzi za achule.
11. Zotsatira
Chitsanzocho chinaneneratu za chule molondola 80.62%.
Kutsiliza
Chabwino, tamaliza ndi phunziro ili. Ngakhale 80.62% siyabwino kwa CNN yaying'ono, ndikukulangizani mwamphamvu kuti musinthe mtundu kapena kuyang'ana ResNet50, Xception, kapena mitundu ina yotsogola kuti mupeze zotsatira zabwino.
Tsopano popeza mwapanga netiweki yanu yoyamba yozindikira zithunzi ku Keras, muyenera kuyesa chitsanzocho kuti muwone momwe magawo amakhudzira magwiridwe ake.
Siyani Mumakonda