Teburin Abubuwan Ciki[Boye][Nuna]
Yana da kwarin gwiwa sanin cewa mun sami nasarar shigar da mutum-mutumi tare da iyawarmu ta asali don koyo ta misali da fahimtar kewayen su. Babban ƙalubalen shine waɗanda ke koyar da kwamfutoci don “gani” kamar yadda mutane za su buƙaci ƙarin lokaci da ƙoƙari sosai.
Koyaya, idan muka yi la'akari da ƙimar amfani da wannan ƙwarewar ke bayarwa a halin yanzu ga ƙungiyoyi da kamfanoni, ƙoƙarin yana da fa'ida. A cikin wannan labarin, zaku koyi game da rarraba hoto, yadda yake aiki, da aiwatar da shi a aikace. Mu fara.
Menene rarraba hoto?
Aikin ciyar da hoto a cikin a neural network kuma samun shi ya fitar da wani nau'i na lakabin wannan hoton ana kiransa da ganewar hoto. Alamar fitarwar hanyar sadarwar zata dace da aji da aka riga aka ayyana.
Ana iya samun azuzuwan da yawa da aka sanya wa hoton, ko kuma ɗaya kawai. Lokacin da aji ɗaya ne, ana amfani da kalmar “gane” akai-akai, yayin da idan akwai azuzuwan da yawa, ana yawan amfani da kalmar “rarrabuwa”.
Gano abu wani yanki ne na rarrabuwar hoto wanda a cikinsa aka gano wasu lokuta na abubuwa na wani aji kamar dabbobi, motoci, ko mutane.
Ta yaya rarraba hoto ke aiki?
Kwamfuta ana nazarin hoto a cikin nau'in pixels. Yana cim ma wannan ta hanyar ɗaukar hoto azaman tarin matrices, girman wanda ƙudurin hoto ya ƙayyade. A taƙaice, rarraba hoto shine nazarin bayanan ƙididdiga ta amfani da algorithm daga mahallin kwamfuta.
Ana aiwatar da rarrabuwar hoto a cikin sarrafa hoto na dijital ta hanyar haɗa pixels zuwa ƙungiyoyin da aka riga aka ƙaddara, ko "aji." Algorithms suna raba hoton zuwa jeri na kyawawan halaye, wanda ke rage nauyi ga mai rarraba na ƙarshe.
Waɗannan halaye suna sanar da mai rarrabawa game da ma'anar hoton da yuwuwar rarrabuwa. Domin sauran hanyoyin da ake bi wajen rarraba hoto sun dogara da shi, hanyar da ake hakowa ita ce mafi mahimmancin lokaci.
The data bayar Algorithm din yana da mahimmanci a cikin rarrabuwar hoto, musamman rarrabawar kulawa. Idan aka kwatanta da mummunan saitin bayanai tare da rashin daidaituwar bayanai dangane da aji da ƙarancin hoto da ingancin rubutu, ingantaccen tsarin rarraba bayanai yana aiki da kyau.
Rarraba hoto ta amfani da Tensorflow & Keras a cikin Python
Za mu yi amfani da CIFAR-10 dataset (wanda ya hada da jiragen sama, jiragen sama, tsuntsaye, da sauran abubuwa 7).
1. Shigar Bukatun
Lambar da ke ƙasa za ta shigar da duk abubuwan da ake buƙata.
2. Shigo da abin dogaro
Yi fayil ɗin train.py a Python. Lambar da ke ƙasa za ta shigo da abubuwan dogaro na Tensorflow da Keras.
3. Ƙaddamar da sigogi
CIFAR-10 ya ƙunshi nau'ikan hoto guda 10 kawai, don haka azuzuwan adadi kawai suna komawa zuwa adadin rukunan don rarrabawa.
4. Load da dataset
Aikin yana amfani da tsarin Tensorflow Datasets don loda saitin bayanai, kuma mun saita tare da bayanai zuwa Gaskiya don samun wasu bayanai game da shi. Kuna iya buga shi don ganin wane fage da ƙimar su, kuma za mu yi amfani da bayanin don dawo da adadin samfuran a cikin tsarin horo da gwaji.
5. Ƙirƙirar samfurin
Yanzu za mu gina yadudduka uku, kowanne ya ƙunshi ConvNets biyu tare da max-pooling da aikin kunnawa ReLU, sannan kuma tsarin haɗin raka'a 1024 cikakke. Idan aka kwatanta da ResNet50 ko Xception, waɗanda sune na'urorin zamani na zamani, wannan na iya zama ƙaramin ƙirar kwatancen.
6. Horar da samfurin
Na yi amfani da Tensorboard don auna daidaito da asara a kowane zamani kuma na samar mana da kyakkyawar nuni bayan shigo da bayanai da samar da samfurin. Guda lambar mai zuwa; dangane da CPU/GPU, horo zai ɗauki mintuna da yawa.
Don amfani da tensorboard, kawai rubuta umarni mai zuwa a cikin tasha ko umarni da sauri a cikin kundin adireshi na yanzu:
Za ku ga cewa asarar tabbatarwa tana raguwa kuma daidaito yana ƙaruwa zuwa kusan 81%. Wannan abin mamaki ne!
Gwajin samfurin
Lokacin da aka gama horo, ana adana samfurin ƙarshe da ma'aunin nauyi a cikin babban fayil ɗin sakamako, yana ba mu damar horar da sau ɗaya kuma mu yi tsinkaya a duk lokacin da muka zaɓa. Bi lambar a cikin sabon fayil ɗin Python mai suna test.py.
7. Shigo da kayan aiki don gwaji
8. Yin kundin adireshi
Yi ƙamus na Python wanda ke fassara kowace ƙima zuwa lakabin da ya dace na dataset:
9. Loading gwajin data & model
Lambar da ke gaba za ta loda bayanan gwajin da samfurin.
10. Kimanta & Hasashen
Lambar da ke biyowa za ta kimanta da yin tsinkaya akan hotunan kwadi.
11. Sakamako
Samfurin ya annabta kwaɗo tare da daidaito 80.62%.
Kammalawa
To, mun gama da wannan darasi. Yayin da 80.62% ba shi da kyau ga ɗan CNN kaɗan, Ina ba ku shawara sosai don canza ƙirar ko duba ResNet50, Xception, ko wasu ƙirar ƙira don ingantacciyar sakamako.
Yanzu da kuka gina cibiyar sadarwar ku ta farko ta gano hoton a cikin Keras, yakamata kuyi gwaji tare da ƙirar don gano yadda sigogi daban-daban ke tasiri aikin sa.
Leave a Reply