S\u0131f\u0131r Maliyetle G\u00fc\u00e7l\u00fc Bir Yapay Zeka Y\u0131\u011f\u0131n\u0131 Olu\u015fturmak: MindsEye x Google AI Mimarisi<\/p>\n
Yapay zeka projeleri geli\u015ftirirken b\u00fct\u00e7e engeline tak\u0131lmak kaderiniz mi? \u00d6zellikle ba\u015flang\u0131\u00e7 a\u015famas\u0131nda, bulut servislerinin ve API maliyetlerinin y\u00fcksekli\u011fi, bir\u00e7ok geli\u015ftiriciyi ve k\u00fc\u00e7\u00fck \u00f6l\u00e7ekli i\u015fletmeyi cayd\u0131rabiliyor. Ancak endi\u015felenmeyin! Bu kapsaml\u0131 rehberde, MindsEye konseptiyle Google AI yeteneklerini harmanlayan, tamamen yerel ve maliyet etkin bir yapay zeka y\u0131\u011f\u0131n\u0131 olu\u015fturman\u0131n inceliklerini ad\u0131m ad\u0131m \u00f6\u011freneceksiniz. S\u0131f\u0131r bulut kredisi veya API b\u00fct\u00e7esiyle dahi, alt\u0131 ayr\u0131 depoyu kullanarak nas\u0131l g\u00fc\u00e7l\u00fc ve \u00f6l\u00e7eklenebilir bir AI altyap\u0131s\u0131 kurabilece\u011finizi ke\u015ffedin. Bu yakla\u015f\u0131m, hem inovasyonunuzu destekleyecek hem de ba\u015flang\u0131\u00e7 maliyetlerinizi minimumda tutman\u0131za yard\u0131mc\u0131 olacak.<\/p>\n
Modern yapay zeka projeleri, genellikle y\u00fcksek performansl\u0131 GPU’lara, geni\u015f veri depolama alanlar\u0131na ve karma\u015f\u0131k bulut altyap\u0131lar\u0131na ihtiya\u00e7 duyar. Ne var ki, bu kaynaklar\u0131n maliyeti, \u00f6zellikle proje ba\u015flang\u0131c\u0131nda veya k\u0131s\u0131tl\u0131 b\u00fct\u00e7elerle \u00e7al\u0131\u015fan ekipler i\u00e7in ciddi bir engel te\u015fkil edebilir. MindsEye x Google AI Stack yakla\u015f\u0131m\u0131, tam da bu noktada devreye girerek, bulut servislerine ba\u011f\u0131ml\u0131l\u0131\u011f\u0131 azalt\u0131p yerel kaynaklar\u0131 optimize ederek g\u00fc\u00e7l\u00fc yapay zeka \u00e7\u00f6z\u00fcmleri geli\u015ftirmenin yolunu a\u00e7ar. Peki, bu kavramlar ne anlama geliyor ve neden bu kadar kritik?<\/p>\n
MindsEye, bu mimaride, bir projenin “alg\u0131” veya “anlama” yetene\u011fini temsil eden bir metafor olarak kullan\u0131lmaktad\u0131r. G\u00f6r\u00fcnt\u00fc i\u015fleme, do\u011fal dil i\u015fleme veya ses analizi gibi alanlarda, karma\u015f\u0131k verileri yerel olarak i\u015fleyerek anlaml\u0131 \u00e7\u0131kar\u0131mlar yapabilen bir sistem hayal edin. Bu, MindsEye’\u0131n temelini olu\u015fturur. \u00d6te yandan, “Google AI Stack” ifadesi, Google’\u0131n yapay zeka alan\u0131ndaki geni\u015f deneyiminden ve geli\u015ftirdi\u011fi modellerden ilham alan yetenek setini ifade eder. Bu, Gemini gibi b\u00fcy\u00fck dil modellerini do\u011frudan kullanmak yerine, onlar\u0131n temel ald\u0131\u011f\u0131 prensipleri veya a\u00e7\u0131k kaynakl\u0131 muadillerini yerel sistemlere entegre etme fikrini i\u00e7erir. Yani, Google’\u0131n sundu\u011fu y\u00fcksek kaliteli AI \u00e7\u00f6z\u00fcmlerinin sundu\u011fu yetenekleri, maliyet etkin ve yerel bir yakla\u015f\u0131mla kendi altyap\u0131m\u0131zda yeniden yaratmaya \u00e7al\u0131\u015f\u0131r\u0131z. Bu bir nevi, Google’\u0131n inovasyonundan ilham alarak kendi \u00f6zelle\u015ftirilmi\u015f ve ba\u011f\u0131ms\u0131z \u00e7\u00f6z\u00fcm\u00fcm\u00fcz\u00fc geli\u015ftirmektir.<\/p>\n
Bu y\u0131\u011f\u0131n\u0131 olu\u015fturan temel bile\u015fenler genellikle a\u00e7\u0131k kaynakl\u0131 k\u00fct\u00fcphaneler (\u00f6rne\u011fin, TensorFlow Lite, PyTorch Mobile, ONNX Runtime, Hugging Face Transformers k\u00fct\u00fcphanesi) ve yerel donan\u0131m \u00fczerinde \u00e7al\u0131\u015fan optimizasyon teknikleridir. Bu yakla\u015f\u0131m\u0131n en b\u00fcy\u00fck avantajlar\u0131ndan biri, veri gizlili\u011fi ve g\u00fcvenli\u011fi konusunda sundu\u011fu ek g\u00fcvencedir. Veriler, buluta g\u00f6nderilmeden yerel olarak i\u015flendi\u011fi i\u00e7in, hassas bilgilerinizin kontrol\u00fc tamamen sizde kal\u0131r. Ayr\u0131ca, internet ba\u011flant\u0131s\u0131 gereksinimi azal\u0131r veya tamamen ortadan kalkar, bu da \u00e7evrimd\u0131\u015f\u0131 \u00e7al\u0131\u015fabilen veya d\u00fc\u015f\u00fck bant geni\u015fli\u011fine sahip ortamlarda kritik uygulamalar geli\u015ftirmek i\u00e7in idealdir. \u00d6rne\u011fin, bir g\u00fcvenlik kameras\u0131 sisteminin, hareket alg\u0131lama ve y\u00fcz tan\u0131ma gibi i\u015flemlerini tamamen yerel bir Raspberry Pi \u00fczerinde yapabildi\u011fini d\u00fc\u015f\u00fcn\u00fcn. Bu, hem anl\u0131k tepki s\u00fcresi sa\u011flar hem de s\u00fcrekli buluta veri g\u00f6nderme maliyetini ortadan kald\u0131r\u0131r. \u00d6zetle, MindsEye x Google AI Stack, esneklik, maliyet etkinli\u011fi, veri gizlili\u011fi ve performans\u0131 bir araya getiren g\u00fc\u00e7l\u00fc bir yerel yapay zeka geli\u015ftirme paradigmas\u0131 sunar.<\/p>\n
Yapay zeka projeleri b\u00fcy\u00fcd\u00fck\u00e7e, bulut servislerinin ve API kullan\u0131mlar\u0131n\u0131n faturas\u0131 da h\u0131zla \u015fi\u015febilir. \u00d6zellikle prototipleme, e\u011fitim veya deneme a\u015famalar\u0131nda, \u00f6ng\u00f6r\u00fclemeyen maliyetler projeleri durma noktas\u0131na getirebilir. Bu durum, \u00f6zellikle k\u0131s\u0131tl\u0131 b\u00fct\u00e7elere sahip bireysel geli\u015ftiriciler, startup’lar ve k\u00fc\u00e7\u00fck i\u015fletmeler i\u00e7in b\u00fcy\u00fck bir problem te\u015fkil eder. Bulut kredi ba\u011f\u0131ml\u0131l\u0131\u011f\u0131, sadece finansal bir y\u00fck olmakla kalmaz, ayn\u0131 zamanda projenin esnekli\u011fini ve gelecekteki \u00f6l\u00e7eklenebilirli\u011fini de k\u0131s\u0131tlayabilir. \u00d6rne\u011fin, b\u00fcy\u00fck bir dil modeli API’si kullanarak her metin \u00e7evirisi veya \u00f6zetleme iste\u011fi i\u00e7in \u00f6deme yapmak, yo\u011fun kullan\u0131mlarda astronomik rakamlara ula\u015fabilir. Bu y\u00fczden, ba\u015flang\u0131\u00e7ta maliyet etkin bir strateji benimsemek, projenin s\u00fcrd\u00fcr\u00fclebilirli\u011fi a\u00e7\u0131s\u0131ndan hayati \u00f6neme sahiptir.<\/p>\n
Peki, bu ba\u011f\u0131ml\u0131l\u0131ktan nas\u0131l kurtuluruz? Cevap, ak\u0131ll\u0131ca yerel kaynak kullan\u0131m\u0131 ve a\u00e7\u0131k kaynakl\u0131 ara\u00e7lar\u0131n g\u00fcc\u00fcnden faydalanmakt\u0131r. \u0130lk ad\u0131m, projenin hangi k\u0131s\u0131mlar\u0131n\u0131n mutlaka bulut gerektirdi\u011fini, hangilerinin yerel olarak \u00e7\u00f6z\u00fclebilece\u011fini net bir \u015fekilde analiz etmektir. \u00c7o\u011fu zaman, veri \u00f6n i\u015fleme, model e\u011fitimi (daha k\u00fc\u00e7\u00fck \u00f6l\u00e7ekli modellerle), \u00e7\u0131kar\u0131m (inference) ve sonu\u00e7lar\u0131n sunumu gibi s\u00fcre\u00e7ler yerel olarak y\u00f6netilebilir. \u00d6rne\u011fin, g\u00f6r\u00fcnt\u00fc tan\u0131ma i\u00e7in, b\u00fcy\u00fck bir bulut tabanl\u0131 API yerine, \u00f6nceden e\u011fitilmi\u015f, optimize edilmi\u015f bir modelin (TensorFlow Lite veya ONNX format\u0131nda) yerel bir cihazda \u00e7al\u0131\u015ft\u0131r\u0131lmas\u0131 hem maliyeti s\u0131f\u0131rlar hem de gecikmeyi \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131r. Bu, \u00f6zellikle kenar bili\u015fim (edge computing) senaryolar\u0131nda, yani verilerin topland\u0131\u011f\u0131 yerde i\u015flendi\u011fi durumlarda paha bi\u00e7ilmezdir.<\/p>\n
\u0130kinci olarak, a\u00e7\u0131k kaynak toplulu\u011funun sundu\u011fu zenginlikten faydalanmak \u00e7ok \u00f6nemlidir. Hugging Face’in Transformers k\u00fct\u00fcphanesi veya spaCy gibi do\u011fal dil i\u015fleme ara\u00e7lar\u0131, bir\u00e7ok dil modelini ve NLP yetene\u011fini yerel olarak \u00e7al\u0131\u015ft\u0131rma imkan\u0131 sunar. G\u00f6r\u00fcnt\u00fc i\u015fleme i\u00e7in OpenCV, makine \u00f6\u011frenimi i\u00e7in Scikit-learn veya daha derin \u00f6\u011frenme modelleri i\u00e7in PyTorch\/TensorFlow’un hafifle\u015ftirilmi\u015f versiyonlar\u0131 bu stratejinin temel ta\u015flar\u0131d\u0131r. Docker ve konteynerizasyon teknolojileri ise, farkl\u0131 yapay zeka bile\u015fenlerini izole edilmi\u015f ortamlarda \u00e7al\u0131\u015ft\u0131rarak ba\u011f\u0131ml\u0131l\u0131k sorunlar\u0131n\u0131 minimize eder ve da\u011f\u0131t\u0131m\u0131 kolayla\u015ft\u0131r\u0131r. Bu sayede, geli\u015ftirme ortam\u0131n\u0131z\u0131 bir kez kurup, farkl\u0131 cihazlarda veya sistemlerde tutarl\u0131 bir \u015fekilde \u00e7al\u0131\u015ft\u0131rabilirsiniz. B\u00f6ylece, projenizin erken a\u015famalar\u0131nda y\u00fczbinlerce dolarl\u0131k bulut faturalar\u0131ndan ka\u00e7\u0131narak, kaynaklar\u0131n\u0131z\u0131 daha stratejik alanlara y\u00f6nlendirebilirsiniz. Bu, sadece bir maliyet avantaj\u0131 de\u011fil, ayn\u0131 zamanda projenize daha fazla kontrol ve esneklik sa\u011flayan bir yakla\u015f\u0131md\u0131r.<\/p>\n
B\u00fcy\u00fck ve karma\u015f\u0131k bir yapay zeka projesini tek bir monolitik kod taban\u0131nda y\u00f6netmek, s\u00fcrd\u00fcr\u00fclebilirlik ve ekip i\u015f birli\u011fi a\u00e7\u0131s\u0131ndan bir\u00e7ok zorluk yarat\u0131r. \u0130\u015fte bu nedenle, MindsEye x Google AI Stack mimarimizde, projemizi alt\u0131 ayr\u0131 depoya (repository) b\u00f6lmek hem mod\u00fclerli\u011fi art\u0131r\u0131r hem de her bir bile\u015fenin ba\u011f\u0131ms\u0131z olarak geli\u015ftirilmesini, test edilmesini ve s\u00fcrd\u00fcr\u00fclmesini kolayla\u015ft\u0131r\u0131r. Bu da\u011f\u0131t\u0131k yap\u0131, ayn\u0131 zamanda farkl\u0131 ekiplerin e\u015f zamanl\u0131 olarak farkl\u0131 mod\u00fcller \u00fczerinde \u00e7al\u0131\u015fabilmesine olanak tan\u0131r ve b\u00f6ylece geli\u015ftirme s\u00fcrecini h\u0131zland\u0131r\u0131r. Peki, bu alt\u0131 depo ne i\u015flev g\u00f6r\u00fcyor ve neden bu ayr\u0131m bu kadar \u00f6nemli?<\/p>\n Bu depo, projenizin “duyusal” organ\u0131 gibidir. G\u00f6r\u00fcnt\u00fcler, metinler, sens\u00f6r verileri veya ses kay\u0131tlar\u0131 gibi ham verileri farkl\u0131 kaynaklardan (yerel dosya sistemi, web kaz\u0131ma, cihaz sens\u00f6rleri) toplar. Daha sonra bu verileri, makine \u00f6\u011frenimi modellerinin anlayabilece\u011fi ve i\u015fleyebilece\u011fi bir formata d\u00f6n\u00fc\u015ft\u00fcr\u00fcr. Bu, g\u00fcr\u00fclt\u00fc azaltma, yeniden boyutland\u0131rma, normalizasyon, etiketleme ve veri art\u0131rma gibi ad\u0131mlar\u0131 i\u00e7erir. Bu ayr\u0131m, veri kaynaklar\u0131n\u0131n de\u011fi\u015fmesi veya yeni \u00f6n i\u015fleme tekniklerinin eklenmesi durumunda, ana model depolar\u0131n\u0131 etkilemeden h\u0131zl\u0131ca adapte olabilmenizi sa\u011flar. Veri kalitesi ve tutarl\u0131l\u0131\u011f\u0131, t\u00fcm yapay zeka projesinin ba\u015far\u0131s\u0131 i\u00e7in temeldir, bu y\u00fczden bu a\u015famay\u0131 ayr\u0131 bir depo olarak ele almak mant\u0131kl\u0131d\u0131r.<\/p>\n<\/li>\n Ad\u0131ndan da anla\u015f\u0131laca\u011f\u0131 gibi, bu depo yapay zeka modellerinizin kalbinin att\u0131\u011f\u0131 yerdir. Se\u00e7ilen mimarilere (\u00f6rne\u011fin, evri\u015fimli sinir a\u011flar\u0131, tekrarlayan sinir a\u011flar\u0131, transformat\u00f6rler) uygun olarak modelleri tan\u0131mlar, e\u011fitir ve hiperparametreleri optimize eder. \u00d6nemli olarak, bu depo ayn\u0131 zamanda modelleri yerel ve kaynak k\u0131s\u0131tl\u0131 ortamlarda \u00e7al\u0131\u015facak \u015fekilde optimize etme (quantization, pruning, knowledge distillation) s\u00fcre\u00e7lerini de i\u00e7erir. \u00d6rne\u011fin, TensorFlow Lite veya ONNX Runtime i\u00e7in model d\u00f6n\u00fc\u015f\u00fcmleri burada ger\u00e7ekle\u015ftirilir. Bu optimizasyonlar, modeli daha k\u00fc\u00e7\u00fck, daha h\u0131zl\u0131 ve daha az bellek t\u00fcketen bir hale getirerek, bulut yerine yerel cihazlarda etkin bir \u015fekilde \u00e7al\u0131\u015fmas\u0131n\u0131 sa\u011flar.<\/p>\n<\/li>\n E\u011fitilmi\u015f ve optimize edilmi\u015f modellerin ger\u00e7ek d\u00fcnya verileri \u00fczerinde tahminler yapmas\u0131n\u0131 sa\u011flayan depodur. Bu, genellikle Flask veya FastAPI gibi hafif bir web \u00e7er\u00e7evesi kullan\u0131larak olu\u015fturulan bir mikro servis veya Python tabanl\u0131 bir CLI arac\u0131 olabilir. Bu depo, \u00f6n i\u015flenmi\u015f verileri al\u0131r, modeli y\u00fckler, \u00e7\u0131kar\u0131m yapar ve sonu\u00e7lar\u0131 uygun bir formatta (\u00f6rne\u011fin, JSON) d\u0131\u015far\u0131 aktar\u0131r. Performans kritik oldu\u011fu i\u00e7in, bu depoda modelin y\u00fcklenme h\u0131z\u0131, \u00e7\u0131kar\u0131m s\u00fcresi ve kaynak t\u00fcketimi \u00f6n plandad\u0131r. Modellerin g\u00fcncellenmesi veya farkl\u0131 modellerin kullan\u0131lmas\u0131 durumunda, Inference Engine deposunu ba\u011f\u0131ms\u0131z olarak g\u00fcncelleyebilir veya \u00f6l\u00e7ekleyebiliriz.<\/p>\n<\/li>\n Son kullan\u0131c\u0131lar\u0131n yapay zeka sisteminizle etkile\u015fim kurdu\u011fu aray\u00fcz\u00fc bar\u0131nd\u0131r\u0131r. Bu bir web uygulamas\u0131 (React, Vue), bir masa\u00fcst\u00fc uygulamas\u0131 (Electron) veya mobil uygulama (Flutter, React Native) olabilir. Bu depo, Inference Engine ile ileti\u015fim kurarak kullan\u0131c\u0131dan gelen girdileri al\u0131r, Inference Engine’e g\u00f6nderir ve oradan gelen sonu\u00e7lar\u0131 anla\u015f\u0131l\u0131r bir \u015fekilde g\u00f6rselle\u015ftirir. Bu ayr\u0131m, aray\u00fcz tasar\u0131mc\u0131lar\u0131n\u0131n ve frontend geli\u015ftiricilerinin backend yapay zeka bile\u015fenlerinden ba\u011f\u0131ms\u0131z olarak \u00e7al\u0131\u015fmas\u0131na olanak tan\u0131r.<\/p>\n<\/li>\n Birden fazla depo taraf\u0131ndan kullan\u0131lan ortak kodlar\u0131, fonksiyonlar\u0131, sabitleri ve temel yap\u0131land\u0131rmalar\u0131 i\u00e7eren bir depodur. \u00d6rne\u011fin, veri modelleme s\u0131n\u0131flar\u0131, \u00f6zel hata i\u015fleme mekanizmalar\u0131, loglama yard\u0131mc\u0131lar\u0131 veya ortak a\u011f istekleri bu depoda yer alabilir. Bu, kod tekrar\u0131n\u0131 \u00f6nler, tutarl\u0131l\u0131\u011f\u0131 art\u0131r\u0131r ve t\u00fcm y\u0131\u011f\u0131n\u0131 daha y\u00f6netilebilir hale getirir. Herhangi bir ortak bile\u015fende yap\u0131lan bir de\u011fi\u015fiklik, tek bir yerde yap\u0131l\u0131p t\u00fcm y\u0131\u011f\u0131na kolayca yay\u0131labilir.<\/p>\n<\/li>\n Bu depo, t\u00fcm bu ba\u011f\u0131ms\u0131z bile\u015fenlerin bir araya getirilip \u00e7al\u0131\u015ft\u0131r\u0131lmas\u0131ndan sorumludur. Docker Compose dosyalar\u0131, Kubernetes yap\u0131land\u0131rmalar\u0131 (ileride buluta ge\u00e7i\u015f durumunda), da\u011f\u0131t\u0131m betikleri (deployment scripts) veya CI\/CD boru hatlar\u0131 bu depoda saklan\u0131r. Bu depo sayesinde, projenin farkl\u0131 ortamlara (geli\u015ftirme, test, \u00fcretim) kolayca da\u011f\u0131t\u0131lmas\u0131 ve ba\u011f\u0131ms\u0131z servislerin bir arada uyumlu bir \u015fekilde \u00e7al\u0131\u015fmas\u0131 sa\u011flan\u0131r. Ba\u015flang\u0131\u00e7ta, bu bir Docker Compose dosyas\u0131 ile t\u00fcm servisleri yerel olarak aya\u011fa kald\u0131rma i\u015flevini g\u00f6rebilir.<\/p>\n<\/ol>\n Bu alt\u0131 depo yap\u0131s\u0131, projenin karma\u015f\u0131kl\u0131\u011f\u0131n\u0131 y\u00f6netmek, her bir bile\u015fenin sorumlulu\u011funu net bir \u015fekilde tan\u0131mlamak ve gelecekteki \u00f6l\u00e7eklenebilirlik i\u00e7in sa\u011flam bir temel olu\u015fturmak ad\u0131na kritik \u00f6neme sahiptir. Her depo, kendi ba\u011f\u0131ms\u0131z ya\u015fam d\u00f6ng\u00fcs\u00fcne sahip olabilir ve farkl\u0131 teknolojilerle (\u00f6rne\u011fin, bir depo Python ile, di\u011feri Node.js ile) geli\u015ftirilebilir, bu da tak\u0131m\u0131n\u0131za b\u00fcy\u00fck esneklik sa\u011flar.<\/p>\n \u015eimdi teoriden prati\u011fe ge\u00e7me zaman\u0131. MindsEye x Google AI Stack’\u0131 yerel olarak, s\u0131f\u0131r bulut kredisi veya API b\u00fct\u00e7esiyle nas\u0131l olu\u015fturabilece\u011fimize dair ad\u0131m ad\u0131m bir rehber sunaca\u011f\u0131z. Bu rehberde, Python tabanl\u0131 bir yakla\u015f\u0131m benimseyecek ve a\u00e7\u0131k kaynak ara\u00e7lar\u0131ndan faydalanaca\u011f\u0131z. Amac\u0131m\u0131z, basit bir g\u00f6r\u00fcnt\u00fc tan\u0131ma servisini bu mimari \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak olsun.<\/p>\n \u00d6ncelikle, t\u00fcm depolar\u0131n kullanaca\u011f\u0131 temel yap\u0131y\u0131 kural\u0131m. Ortak ara\u00e7lar ve da\u011f\u0131t\u0131m betikleri i\u00e7in iki ayr\u0131 depo olu\u015fturaca\u011f\u0131z. \u0130lk olarak, bir <\/code><\/p>\n Daha sonra, bu servisleri aya\u011fa kald\u0131racak bir <\/code><\/p>\n Bu depo, raw g\u00f6r\u00fcnt\u00fcleri alacak ve modelimizin bekledi\u011fi formata d\u00f6n\u00fc\u015ft\u00fcrecek bir mikro servis i\u00e7erecek. FastAPI ve Pillow k\u00fct\u00fcphanelerini kullanaca\u011f\u0131z.<\/p>\n <\/code><\/p>\n <\/code><\/p>\n Bu depo, basit bir g\u00f6r\u00fcnt\u00fc s\u0131n\u0131fland\u0131rma modeli e\u011fitecek ve onu TensorFlow Lite format\u0131na d\u00f6n\u00fc\u015ft\u00fcrecektir. Model dosyas\u0131 daha sonra <\/code><\/p>\n <\/code> Bu depo, optimize edilmi\u015f TFLite modelini y\u00fckleyecek ve \u00f6n i\u015flenmi\u015f verileri al\u0131p tahminler yapacak bir FastAPI servisi i\u00e7erecek.<\/p>\n <\/code><\/p>\n <\/code><\/p>\n Basit bir web aray\u00fcz\u00fc olu\u015fturmak i\u00e7in JavaScript (Fetch API) ve HTML kullanaca\u011f\u0131z. Bu depo, \u00f6n i\u015fleme servisine bir g\u00f6r\u00fcnt\u00fc g\u00f6nderecek, ard\u0131ndan gelen veriyi \u00e7\u0131kar\u0131m servisine aktaracak ve sonu\u00e7lar\u0131 g\u00f6sterecektir.<\/p>\n
\n Uzman \u0130pucu: Bulut kredilerinden tamamen ka\u00e7\u0131nmak yerine, bunlar\u0131 stratejik olarak kullanmay\u0131 d\u00fc\u015f\u00fcn\u00fcn. \u00d6rne\u011fin, ba\u015flang\u0131\u00e7ta \u00e7ok b\u00fcy\u00fck modelleri e\u011fitmek i\u00e7in k\u0131sa s\u00fcreli\u011fine bulut GPU’lar\u0131ndan faydalan\u0131p, ard\u0131ndan bu modelleri optimize ederek yerel \u00e7\u0131kar\u0131m i\u00e7in kullanabilirsiniz. Bu, “hibrit” bir yakla\u015f\u0131m olup, her iki d\u00fcnyan\u0131n da en iyi yanlar\u0131n\u0131 birle\u015ftirir.
\n<\/span><\/p>\nMimariyi Tasarlamak: 6 Ayr\u0131 Depo Neden Gerekli ve Ne \u0130\u015flev G\u00f6r\u00fcyor?<\/h2>\n
\n
Ad\u0131m Ad\u0131m Uygulama: MindsEye x Google AI Stack Nas\u0131l Olu\u015fturulur?<\/h2>\n
1. Ortak Altyap\u0131 Haz\u0131rl\u0131\u011f\u0131 (Ortak Yard\u0131mc\u0131lar ve Da\u011f\u0131t\u0131m Depolar\u0131)<\/h3>\n
shared_utilities<\/code> deposu ile ba\u015fl\u0131yoruz. Bu depo, loglama ayarlar\u0131n\u0131 ve baz\u0131 temel veri modellerini i\u00e7erebilir.<\/p>\n\n# shared_utilities\/logger_config.py\nimport logging\nimport sys\n\ndef setup_logger(name, level=logging.INFO):\n formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n handler = logging.StreamHandler(sys.stdout)\n handler.setFormatter(formatter)\n logger = logging.getLogger(name)\n logger.setLevel(level)\n if not logger.handlers:\n logger.addHandler(handler)\n return logger\n\n# shared_utilities\/models.py\nfrom pydantic import BaseModel\nfrom typing import List, Dict, Any\n\nclass ImagePrediction(BaseModel):\n label: str\n confidence: float\n\nclass InferenceResult(BaseModel):\n predictions: List[ImagePrediction]\n processed_time_ms: float\n image_id: str\n<\/pre>\ndeployment<\/code> deposu olu\u015fturaca\u011f\u0131z. Ba\u015flang\u0131\u00e7 i\u00e7in bir docker-compose.yml<\/code> dosyas\u0131 idealdir. Bu dosya, ilerleyen ad\u0131mlarda olu\u015fturaca\u011f\u0131m\u0131z servisleri bir arada \u00e7al\u0131\u015ft\u0131racakt\u0131r.<\/p>\n\n# deployment\/docker-compose.yml\nversion: '3.8'\nservices:\n inference-engine:\n build: ..\/inference_engine\n ports:\n - \"8000:8000\"\n volumes:\n - .\/models:\/app\/models # Modelleri Inference Engine'e mount ediyoruz\n environment:\n - MODEL_PATH=\/app\/models\/optimized_model.tflite\n depends_on:\n - data-preprocessing\n data-preprocessing:\n build: ..\/data_preprocessing\n ports:\n - \"8001:8001\"\n ui-application:\n build: ..\/ui_application\n ports:\n - \"3000:3000\"\n environment:\n - INFERENCE_API_URL=http:\/\/inference-engine:8000\n<\/pre>\n2. Veri Al\u0131m\u0131 ve \u00d6n \u0130\u015fleme Deposu Olu\u015fturma (
data_preprocessing<\/code>)<\/h3>\n\n# data_preprocessing\/main.py\nfrom fastapi import FastAPI, UploadFile, File\nfrom PIL import Image\nimport io\nimport numpy as np\nfrom shared_utilities.logger_config import setup_logger # shared_utilities'dan loglay\u0131c\u0131y\u0131 i\u00e7e aktar\nimport uvicorn\n\nlogger = setup_logger(__name__)\napp = FastAPI()\n\n@app.post(\"\/preprocess_image\/\")\nasync def preprocess_image(file: UploadFile = File(...)):\n try:\n image_bytes = await file.read()\n image = Image.open(io.BytesIO(image_bytes))\n \n # \u00d6rnek \u00f6n i\u015fleme: Boyutland\u0131rma ve gri tonlama\n image = image.resize((224, 224)) # Modelin bekledi\u011fi boyuta getir\n image = image.convert(\"RGB\") # RGB'ye d\u00f6n\u00fc\u015ft\u00fcr (e\u011fer de\u011filse)\n \n # NumPy dizisine \u00e7evir ve normalle\u015ftir\n img_array = np.array(image).astype(np.float32)\n img_array = img_array \/ 255.0 # 0-1 aras\u0131na normalle\u015ftir\n img_array = np.expand_dims(img_array, axis=0) # Batch boyutu ekle (1, 224, 224, 3)\n\n logger.info(f\"Image {file.filename} preprocessed successfully. Shape: {img_array.shape}\")\n return {\"message\": \"Image preprocessed\", \"shape\": img_array.shape.tolist(), \"data\": img_array.tobytes().hex()}\n except Exception as e:\n logger.error(f\"Error during image preprocessing: {e}\")\n return {\"message\": \"Error processing image\", \"error\": str(e)}, 500\n\nif __name__ == \"__main__\":\n uvicorn.run(app, host=\"0.0.0.0\", port=8001)\n\n<\/pre>\n\n# data_preprocessing\/Dockerfile\nFROM python:3.9-slim-buster\nWORKDIR \/app\nCOPY requirements.txt .\nRUN pip install --no-cache-dir -r requirements.txt\nCOPY . .\n# shared_utilities deposunu buraya kopyalamam\u0131z gerekiyor\nCOPY ..\/shared_utilities \/app\/shared_utilities\nCMD [\"uvicorn\", \"main:app\", \"--host\", \"0.0.0.0\", \"--port\", \"8001\"]\n\n# data_preprocessing\/requirements.txt\nfastapi\nuvicorn\nPillow\nnumpy\npydantic # Shared utilities'daki modeller i\u00e7in\n<\/pre>\n3. Model E\u011fitimi ve Optimizasyonu Deposu Olu\u015fturma (
model_training_optimization<\/code>)<\/h3>\ndeployment\/models<\/code> klas\u00f6r\u00fcne kopyalanacakt\u0131r.<\/p>\n\n# model_training_optimization\/train_and_optimize.py\nimport tensorflow as tf\nfrom tensorflow import keras\nimport numpy as np\nfrom shared_utilities.logger_config import setup_logger # shared_utilities'dan loglay\u0131c\u0131y\u0131 i\u00e7e aktar\nimport os\n\nlogger = setup_logger(__name__)\n\ndef create_and_train_model():\n # Basit bir model olu\u015ftural\u0131m: MobileNetV2 \u00fczerine kurulu\n base_model = keras.applications.MobileNetV2(input_shape=(224, 224, 3), include_top=False, weights='imagenet')\n base_model.trainable = False # \u00d6nceden e\u011fitilmi\u015f katmanlar\u0131 dondur\n\n inputs = keras.Input(shape=(224, 224, 3))\n x = base_model(inputs, training=False)\n x = keras.layers.GlobalAveragePooling2D()(x)\n outputs = keras.layers.Dense(10, activation='softmax')(x) # \u00d6rnek olarak 10 s\u0131n\u0131f\n \n model = keras.Model(inputs, outputs)\n model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n \n # Basit bir dummy veri seti olu\u015ftural\u0131m (ger\u00e7ek uygulamada kendi verinizi kullan\u0131n)\n dummy_x = np.random.rand(100, 224, 224, 3).astype(np.float32)\n dummy_y = np.random.randint(0, 10, 100)\n\n logger.info(\"Starting model training with dummy data...\")\n model.fit(dummy_x, dummy_y, epochs=3, verbose=1)\n logger.info(\"Model training finished.\")\n \n return model\n\ndef optimize_model(model, output_path=\"deployment\/models\/optimized_model.tflite\"):\n converter = tf.lite.TFLiteConverter.from_keras_model(model)\n # Varsay\u0131lan optimizasyon (quantization)\n converter.optimizations = [tf.lite.Optimize.DEFAULT]\n tflite_model = converter.convert()\n\n os.makedirs(os.path.dirname(output_path), exist_ok=True)\n with open(output_path, 'wb') as f:\n f.write(tflite_model)\n logger.info(f\"Optimized TFLite model saved to {output_path}\")\n\nif __name__ == \"__main__\":\n trained_model = create_and_train_model()\n optimize_model(trained_model)\n\n<\/pre>\n\n# model_training_optimization\/requirements.txt\ntensorflow==2.x # Y\u00fckl\u00fc TensorFlow s\u00fcr\u00fcm\u00fcn\u00fcze g\u00f6re ayarlay\u0131n\nnumpy\n<\/pre>\n
\n
\n Uzman \u0130pucu: Model e\u011fitiminde GPU kullanmak i\u00e7in TensorFlow kurulumunuzu ve Docker konfig\u00fcrasyonunuzu (nvidia-docker) buna g\u00f6re ayarlaman\u0131z gerekir. Ancak ba\u015flang\u0131\u00e7 i\u00e7in CPU'da k\u00fc\u00e7\u00fck modellerle denemeler yapmak maliyet a\u00e7\u0131s\u0131ndan daha uygun olacakt\u0131r.
\n<\/span><\/p>\n4. \u00c7\u0131kar\u0131m Motoru Deposu Olu\u015fturma (
inference_engine<\/code>)<\/h3>\n\n# inference_engine\/main.py\nfrom fastapi import FastAPI, HTTPException\nimport tensorflow as tf\nimport numpy as np\nimport os\nimport time\nfrom shared_utilities.logger_config import setup_logger\nfrom shared_utilities.models import InferenceResult, ImagePrediction # Payla\u015f\u0131lan modelleri i\u00e7e aktar\nimport uvicorn\n\nlogger = setup_logger(__name__)\napp = FastAPI()\n\ninterpreter = None\ninput_details = None\noutput_details = None\n\n# \u00d6rnek etiketler (modelin e\u011fitildi\u011fi s\u0131n\u0131flara g\u00f6re g\u00fcncelleyin)\nLABELS = [f\"class_{i}\" for i in range(10)] \n\n@app.on_event(\"startup\")\nasync def load_model():\n global interpreter, input_details, output_details\n model_path = os.getenv(\"MODEL_PATH\", \"\/app\/models\/optimized_model.tflite\")\n if not os.path.exists(model_path):\n logger.error(f\"Model file not found at {model_path}\")\n raise FileNotFoundError(f\"Model file not found at {model_path}\")\n\n logger.info(f\"Loading TFLite model from {model_path}\")\n interpreter = tf.lite.Interpreter(model_path=model_path)\n interpreter.allocate_tensors()\n input_details = interpreter.get_input_details()\n output_details = interpreter.get_output_details()\n logger.info(\"TFLite model loaded successfully.\")\n\n@app.post(\"\/predict\/\", response_model=InferenceResult)\nasync def predict_image(image_data: dict):\n if interpreter is None:\n raise HTTPException(status_code=500, detail=\"Model not loaded. Server starting up.\")\n\n try:\n # \u00d6n i\u015flenmi\u015f veriyi hex string'den numpy dizisine \u00e7evir\n # image_data.get(\"data\") preprocessor'dan gelen hex string olmal\u0131\n img_array_bytes = bytes.fromhex(image_data[\"data\"])\n img_array = np.frombuffer(img_array_bytes, dtype=np.float32).reshape(image_data[\"shape\"])\n\n input_tensor = input_details[0]['index']\n output_tensor = output_details[0]['index']\n\n interpreter.set_tensor(input_tensor, img_array)\n \n start_time = time.time()\n interpreter.invoke()\n end_time = time.time()\n \n predictions = interpreter.get_tensor(output_tensor)[0]\n \n processed_time_ms = (end_time - start_time) * 1000\n logger.info(f\"Inference completed in {processed_time_ms:.2f} ms.\")\n\n # Tahminleri ImagePrediction listesine \u00e7evir\n results = []\n for i, score in enumerate(predictions):\n results.append(ImagePrediction(label=LABELS[i], confidence=float(score)))\n \n return InferenceResult(\n predictions=sorted(results, key=lambda x: x.confidence, reverse=True),\n processed_time_ms=processed_time_ms,\n image_id=image_data.get(\"image_id\", \"N\/A\")\n )\n\n except Exception as e:\n logger.error(f\"Error during inference: {e}\")\n raise HTTPException(status_code=500, detail=f\"Inference error: {e}\")\n\nif __name__ == \"__main__\":\n uvicorn.run(app, host=\"0.0.0.0\", port=8000)\n\n<\/pre>\n\n# inference_engine\/Dockerfile\nFROM python:3.9-slim-buster\nWORKDIR \/app\nCOPY requirements.txt .\nRUN pip install --no-cache-dir -r requirements.txt\nCOPY . .\nCOPY ..\/shared_utilities \/app\/shared_utilities # shared_utilities deposunu buraya da kopyal\u0131yoruz\nCMD [\"uvicorn\", \"main:app\", \"--host\", \"0.0.0.0\", \"--port\", \"8000\"]\n\n# inference_engine\/requirements.txt\nfastapi\nuvicorn\ntensorflow # TFLite interpreter i\u00e7in\nnumpy\npydantic\n<\/pre>\n5. Kullan\u0131c\u0131 Aray\u00fcz\u00fc\/Uygulama Deposu Olu\u015fturma (
ui_application<\/code>)<\/h3>\n\n\n\n MindsEye Yerel G\u00f6r\u00fcnt\u00fc S\u0131n\u0131fland\u0131r\u0131c\u0131<\/h1>\n \n