Hasil Pencarian  ::  Simpan CSV :: Kembali

"ABSTRAKDensely Connected Convolutional Networks (DenseNet) merupakan salah satumodel arsitektur Deep Learning yang menghubungkan setiap layer beserta feature-maps ke seluruh layer berikutnya, sehingga layer berikutnya menerima inputfeature-maps dari seluruh layer sebelumnya. Karena padatnya arsitektur DenseNetmeyebabkan komputasi model memerlukan waktu lama dan pemakaian memoryGPU yang besar. Penelitian ini mengembangkan metode optimisasi DenseNetmenggunakan batching strategy yang bertujuan untuk mengatasi permasalahanDenseNet dalam hal percepatan komputasi dan penghematan ruang memory GPU.Batching strategy adalah metode yang digunakan dalam Stochastic GradientDescent (SGD) dimana metode tersebut menerapkan metode dinamik batchingdengan inisialisasi awal menggunakan ukuran batch kecil dan ditingkatkanukurannya secara adaptif selama training hingga sampai ukuran batch besar agarterjadi peningkatan paralelisasi komputasi untuk mempercepat waktu pelatihan.Metode batching strategy juga dilengkapi dengan manajemen memory GPUmenggunakan metode gradient accumulation. Dari hasil percobaan dan pengujianterhadap metode tersebut dihasilkan peningkatan kecepatan waktu pelatihan hingga1,7x pada dataset CIFAR-10 dan 1,5x pada dataset CIFAR-100 serta dapatmeningkatkan akurasi DenseNet. Manajemen memory yang digunakan dapatmenghemat memory GPU hingga 30% jika dibandingkan dengan native DenseNet.Dataset yang digunakan menggunakan CIFAR-10 dan CIFAR-100 datasets.Penerapan metode batching strategy tersebut terbukti dapat menghasilkanpercepatan dan penghematan ruang memory GPU.

---

ABSTRACTDensely Connected Convolutional Networks (DenseNet) is one of the DeepLearning architecture models that connect each layer and feature maps to allsubsequent layers so that the next layer receives input feature maps from allprevious layers. Because of its DenseNet architecture, computational modelsrequire a long time and use large GPU memory. This research develops theDenseNet optimization method using a batching strategy that aims to overcome theDenseNet problem in terms of accelerating computing time and saving GPUmemory. Batching strategy is a method used in Stochastic Gradient Descent (SGD)where the technique applies dynamic batching approach with initial initializationusing small batch sizes and adaptively increased size during training to large batchsizes so that there is an increase in computational parallelization to speed up trainingtime. The batching strategy method is also equipped with GPU memorymanagement using the gradient accumulation method. From the results ofexperiments and testing of these methods resulted in an increase in training timespeed of up to 1.7x on the CIFAR-10 dataset and 1.5x on the CIFAR-100 datasetand can improve DenseNet accuracy. Memory management used can save GPUmemory up to 30% when compared to native DenseNet. The dataset used usesCIFAR-10 and CIFAR-100 datasets. The application of the batching strategymethod is proven to be able to produce acceleration and saving of GPU memory."

"This book shows you how ARC works and how best to incorporate it into your applications. Grand Central Dispatch (GCD) and blocks are key to developing great apps, allowing you to control threads for maximum performance.If for you, multithreading is an unsolved mystery and ARC is unexplored territory, then this is the book you'll need to make these concepts clear and send you on your way to becoming a master iOS and OS X developer. What are blocks? How are they used with GCD? Multithreading with GCD. Managing objects with ARC."

"Practical load balancing presents an entire analytical framework to increase performance not just of one machine, but of your entire infrastructure.Practical load balancing starts by introducing key concepts and the tools you'll need to tackle your load-balancing issues. You'll travel through the IP layers and learn how they can create increased network traffic for you. You'll see how to account for persistence and state, and how you can judge the performance of scheduling algorithms.You'll then learn how to avoid performance degradation and any risk of the sudden disappearance of a service on a server. If you're concerned with running your load balancer for an entire network, you'll find out how to set up your network topography, and condense each topographical variety into recipes that will serve you in different situations. You'll also learn about individual servers, and load balancers that can perform cookie insertion or improve your SSL throughput. You'll also explore load balancing in the modern context of the cloud. While load balancers need to be configured for high availability once the conditions on the network have been created, modern load balancing has found its way into the cloud, where good balancing is vital for the very functioning of the cloud, and where IPv6 is becoming ever more important."