Work

Prestatieverbeterende technieken voor Tuckertensorontbindingen in dataverwerking

Year: 2023
Type: dissertation
Abstract: As datasets grow in size, so has their dimensionality in many cases. Numerical data in the form of arrays with 3 or more dimensions, i.e. tensors, are commonly found in hyperspectral imaging, diffusion tensor imaging, simulations, X-ray scans, and so on. Even relational databases can be represented using sparse tensors. As a result of this trend, tensor decompositions were developed to process such multidimensional numerical data. Among these, the Tucker decomposition is a very popular tool for data compression and preprocessing tensors for further analysis. As the field of numerical multilinear algebra is still relatively young, many problems related to this decomposition are only solved suboptimally or not at all, especially concerning their efficient algorithmic implementation. The purpose of this thesis is to develop techniques that enhance the performance of existing solutions to some of the aforementioned problems, both in terms of algorithmic and implementation improvements. We empirically verify the performance of our proposed techniques using experiments with real-world datasets. Within the context of these objectives, two major problems are studied. In the first part, we focus on the problem of Tucker-based data compression of tensors to the byte level. To this end, we build upon the pipeline of the state-of-the-art TTHRESH compressor to develop our own novel compression software, the Advanced Tucker Compressor (ATC). We propose a series of algorithmic changes to improve speed, memory usage, compression rate and error control. First, we describe a hybrid truncation scheme that combines rank and bit plane truncation while maintaining error control. Second, we parallelize the quantization and encoding phase, which represents a major part of the compression pipeline. Third, we drastically improve error control by taking the dequantization correction into account during compression. Moreover, we propose many other minor optimizations and describe several important implementation aspects. Our experiments show that ATC maintains state-of-the-art compression rates while achieving average speed-ups of 2.2 to 3.5, halving memory usage and providing 25 times more precise error control. Compared to non-tensor-based compressors, ATC also reaches much higher compression rates in the high-compression domain. Finally, ATC is also developed as a high-quality parallelized software library, with multiple interfaces, broad support for data types, different types of I/O and extensive documentation. In the second part, we explore the problem of computing sparse tensor-times-matrix chains (TTMCs), a key kernel during the calculation of Tucker decompositions. We describe the intermediate blowup problem faced by naive methods and discuss several solutions, leading up to the state-of-the-art algorithm based on the Compressed Sparse Fiber (CSF) sparse tensor format. We propose three derived algorithms based on the principle of all-accumulation to further improve the performance of this method without using redundant tensor representations. Within this setting, our method halves the cost of leave-one-out TTMCs and also reduces the amount of required operations as well the runtime of full TTMCs by 70% and 38% respectively. Finally, we also observe massive reductions in memory usage of 23x to 75x compared to the reference implementation of the existing CSF-based work.
Source: Lirias
Language: English
Open Access status: green
OpenAlex ID: W7070759208
Cites: 64
Cited by:
Related to: 10