-
Memory Efficient Diffusion Probabilistic Models via Patch-based Generation
Authors:
Shinei Arakawa,
Hideki Tsunashima,
Daichi Horita,
Keitaro Tanaka,
Shigeo Morishima
Abstract:
Diffusion probabilistic models have been successful in generating high-quality and diverse images. However, traditional models, whose input and output are high-resolution images, suffer from excessive memory requirements, making them less practical for edge devices. Previous approaches for generative adversarial networks proposed a patch-based method that uses positional encoding and global conten…
▽ More
Diffusion probabilistic models have been successful in generating high-quality and diverse images. However, traditional models, whose input and output are high-resolution images, suffer from excessive memory requirements, making them less practical for edge devices. Previous approaches for generative adversarial networks proposed a patch-based method that uses positional encoding and global content information. Nevertheless, designing a patch-based approach for diffusion probabilistic models is non-trivial. In this paper, we resent a diffusion probabilistic model that generates images on a patch-by-patch basis. We propose two conditioning methods for a patch-based generation. First, we propose position-wise conditioning using one-hot representation to ensure patches are in proper positions. Second, we propose Global Content Conditioning (GCC) to ensure patches have coherent content when concatenated together. We evaluate our model qualitatively and quantitatively on CelebA and LSUN bedroom datasets and demonstrate a moderate trade-off between maximum memory consumption and generated image quality. Specifically, when an entire image is divided into 2 x 2 patches, our proposed approach can reduce the maximum memory consumption by half while maintaining comparable image quality.
△ Less
Submitted 14 April, 2023;
originally announced April 2023.
-
Serializable HTAP with Abort-/Wait-free Snapshot Read
Authors:
Takamitsu Shioi,
Takashi Kambayashi,
Suguru Arakawa,
Ryoji Kurosawa,
Satoshi Hikida,
Haruo Yokota
Abstract:
Concurrency Control (CC) ensuring consistency of updated data is an essential element of OLTP systems. Recently, hybrid transactional/analytical processing (HTAP) systems developed for executing OLTP and OLAP have attracted much attention. The OLAP side CC domain has been isolated from OLTP's CC and in many cases has been achieved by snapshot isolation (SI) to establish HTAP systems. Although high…
▽ More
Concurrency Control (CC) ensuring consistency of updated data is an essential element of OLTP systems. Recently, hybrid transactional/analytical processing (HTAP) systems developed for executing OLTP and OLAP have attracted much attention. The OLAP side CC domain has been isolated from OLTP's CC and in many cases has been achieved by snapshot isolation (SI) to establish HTAP systems. Although higher isolation level is ideal, considering OLAP read-only transactions in the context of OLTP scheduling achieving serializability forces aborts/waits and would be a potential performance problem. Furthermore, executing OLAP without affecting OLTP as much as possible is needed for HTAP systems.
The aim of this study was serializability without additional aborts/waits. We propose read safe snapshot (RSS) using multiversion CC (MVCC) theory and introduce the RSS construction algorithm utilizing serializable snapshot isolation (SSI). For serializability of HTAP systems, our model makes use of multiversion and allows more schedules with read operations whose corresponding write operations do not participate in the dependency cycles. Furthermore, we implemented the algorithm practically in an open-source database system that offers SSI. Our algorithm was integrated into two types of architecture as HTAP systems called as unified (single-node) or decoupled (multinode) storage architecture. We evaluate the performance and abort rate of the single-node architecture where SSI is applicable. The multi-node architecture was investigated for examining the performance overhead applying our algorithm.
△ Less
Submitted 21 April, 2022; v1 submitted 19 January, 2022;
originally announced January 2022.
-
A Highly Tunable Virtual Topology Controller
Authors:
Y. Sinan Hanay,
Shinichi Arakawa,
Masayuki Murata
Abstract:
Much research in the last two decades has focused on Virtual Topology Reconfiguration (VTR) problem. However, most of the proposed methods either has low controllability, or the analysis of a control parameter is limited to empirical analysis. In this paper, we present a highly tunable Virtual Topology (VT) controller. First, we analyze the controllability of two previously proposed VTR algorithms…
▽ More
Much research in the last two decades has focused on Virtual Topology Reconfiguration (VTR) problem. However, most of the proposed methods either has low controllability, or the analysis of a control parameter is limited to empirical analysis. In this paper, we present a highly tunable Virtual Topology (VT) controller. First, we analyze the controllability of two previously proposed VTR algorithms: a heuristic method and a neural networks based method. Then we present insights on how to transform these VTR methods to their tunable versions. To benefit from the controllability, an optimality analysis of the control parameter is needed. In the second part of the paper, through a probabilistic analysis we find an optimal parameter for the neural network based method. We validated our analysis through simulations. We propose this highly tunable method as a new VTR algorithm.
△ Less
Submitted 22 January, 2015;
originally announced January 2015.
