PyTorchKR Papers

최신 논문

총 5,576편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2202.11921공식 코드 · Papers with Code

Auto-scaling Vision Transformers without Training

This work targets automated designing and scaling of Vision Transformers (ViTs). The motivation comes from two pain spots: 1) the lack of efficient and principled methods for designing and scaling ViTs; 2) the tremendous computational cost of training ViT that is much heavier than its convolution counterpart. To tackle these issues, we propose As-ViT, an auto-scaling framework for ViTs without training, which automatically discovers and scales up ViTs in an efficient and principled manner. Specifically, we first design a "seed" ViT topology by leveraging a training-free search process. This extremely fast search is fulfilled by a comprehensive study of ViT's network complexity, yielding a strong Kendall-tau correlation with ground-truth accuracies. Second, starting from the "seed" topology, we automate the scaling rule for ViTs by growing widths/depths to different ViT layers. This results in a series of architectures with different numbers of parameters in a single run. Finally, based on the observation that ViTs can tolerate coarse tokenization in early training stages, we propose a progressive tokenization strategy to train ViTs faster and cheaper. As a unified framework, As-ViT achieves strong performance on classification (83.5% top1 on ImageNet-1k) and detection (52.7% mAP on COCO) without any manual crafting nor scaling of ViT architectures: the end-to-end model design and scaling process cost only 12 hours on one V100 GPU. Our code is available at https://github.com/VITA-Group/AsViT.

arXiv:2202.10533

Dynamic Sampling Rate: Harnessing Frame Coherence in Graphics Applications for Energy-Efficient GPUs

In real-time rendering, a 3D scene is modelled with meshes of triangles that the GPU projects to the screen. They are discretized by sampling each triangle at regular space intervals to generate fragments which are then added texture and lighting effects by a shader program. Realistic scenes require detailed geometric models, complex shaders, high-resolution displays and high screen refreshing rates, which all come at a great compute time and energy cost. This cost is often dominated by the fragment shader, which runs for each sampled fragment. Conventional GPUs sample the triangles once per pixel, however, there are many screen regions containing low variation that produce identical fragments and could be sampled at lower than pixel-rate with no loss in quality. Additionally, as temporal frame coherence makes consecutive frames very similar, such variations are usually maintained from frame to frame. This work proposes Dynamic Sampling Rate (DSR), a novel hardware mechanism to reduce redundancy and improve the energy efficiency in graphics applications. DSR analyzes the spatial frequencies of the scene once it has been rendered. Then, it leverages the temporal coherence in consecutive frames to decide, for each region of the screen, the lowest sampling rate to employ in the next frame that maintains image quality. We evaluate the performance of a state-of-the-art mobile GPU architecture extended with DSR for a wide variety of applications. Experimental results show that DSR is able to remove most of the redundancy inherent in the color computations at fragment granularity, which brings average speedups of 1.68x and energy savings of 40%.

arXiv:2112.10408공식 코드 · Papers with Code

Efficient Wind Speed Nowcasting with GPU-Accelerated Nearest Neighbors Algorithm

This paper proposes a simple yet efficient high-altitude wind nowcasting pipeline. It processes efficiently a vast amount of live data recorded by airplanes over the whole airspace and reconstructs the wind field with good accuracy. It creates a unique context for each point in the dataset and then extrapolates from it. As creating such context is computationally intensive, this paper proposes a novel algorithm that reduces the time and memory cost by efficiently fetching nearest neighbors in a data set whose elements are organized along smooth trajectories that can be approximated with piece-wise linear structures. We introduce an efficient and exact strategy implemented through algebraic tensorial operations, which is well-suited to modern GPU-based computing infrastructure. This method employs a scalable Euclidean metric and allows masking data points along one dimension. When applied, this method is more efficient than plain Euclidean k-NN and other well-known data selection methods such as KDTrees and provides a several-fold speedup. We provide an implementation in PyTorch and a novel data set to allow the replication of empirical results.

arXiv:2202.11001

Multi-Objective Dual Simplex-Mesh Based Deformable Image Registration for 3D Medical Images -- Proof of Concept

Reliably and physically accurately transferring information between images through deformable image registration with large anatomical differences is an open challenge in medical image analysis. Most existing methods have two key shortcomings: first, they require extensive up-front parameter tuning to each specific registration problem, and second, they have difficulty capturing large deformations and content mismatches between images. There have however been developments that have laid the foundation for potential solutions to both shortcomings. Towards the first shortcoming, a multi-objective optimization approach using the Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (RV-GOMEA) has been shown to be capable of producing a diverse set of registrations for 2D images in one run of the algorithm, representing different trade-offs between conflicting objectives in the registration problem. This allows the user to select a registration afterwards and removes the need for up-front tuning. Towards the second shortcoming, a dual-dynamic grid transformation model has proven effective at capturing large differences in 2D images. These two developments have recently been accelerated through GPU parallelization, delivering large speed-ups. Based on this accelerated version, it is now possible to extend the approach to 3D images. Concordantly, this work introduces the first method for multi-objective 3D deformable image registration, using a 3D dual-dynamic grid transformation model based on simplex meshes while still supporting the incorporation of annotated guidance information and multi-resolution schemes. Our proof-of-concept prototype shows promising results on synthetic and clinical 3D registration problems, forming the foundation for a new, insightful method that can include bio-mechanical properties in the registration.

arXiv:2202.10100

Enabling On-Device Smartphone GPU based Training: Lessons Learned

Deep Learning (DL) has shown impressive performance in many mobile applications. Most existing works have focused on reducing the computational and resource overheads of running Deep Neural Networks (DNN) inference on resource-constrained mobile devices. However, the other aspect of DNN operations, i.e. training (forward and backward passes) on smartphone GPUs, has received little attention thus far. To this end, we conduct an initial analysis to examine the feasibility of on-device training on smartphones using mobile GPUs. We first employ the open-source mobile DL framework (MNN) and its OpenCL backend for running compute kernels on GPUs. Next, we observed that training on CPUs is much faster than on GPUs and identified two possible bottlenecks related to this observation: (i) computation and (ii) memory bottlenecks. To solve the computation bottleneck, we optimize the OpenCL backend's kernels, showing 2x improvements (40-70 GFLOPs) over CPUs (15-30 GFLOPs) on the Snapdragon 8 series processors. However, we find that the full DNN training is still much slower on GPUs than on CPUs, indicating that memory bottleneck plays a significant role in the lower performance of GPU over CPU. The data movement takes almost 91% of training time due to the low bandwidth. Lastly, based on the findings and failures during our investigation, we present limitations and practical guidelines for future directions.

arXiv:2202.09518공식 코드 · Papers with Code

Distributed Out-of-Memory NMF on CPU/GPU Architectures

We propose an efficient distributed out-of-memory implementation of the Non-negative Matrix Factorization (NMF) algorithm for heterogeneous high-performance-computing (HPC) systems. The proposed implementation is based on prior work on NMFk, which can perform automatic model selection and extract latent variables and patterns from data. In this work, we extend NMFk by adding support for dense and sparse matrix operation on multi-node, multi-GPU systems. The resulting algorithm is optimized for out-of-memory (OOM) problems where the memory required to factorize a given matrix is greater than the available GPU memory. Memory complexity is reduced by batching/tiling strategies, and sparse and dense matrix operations are significantly accelerated with GPU cores (or tensor cores when available). Input/Output (I/O) latency associated with batch copies between host and device is hidden using CUDA streams to overlap data transfers and compute asynchronously, and latency associated with collective communications (both intra-node and inter-node) is reduced using optimized NVIDIA Collective Communication Library NCCL based communicators. Benchmark results show significant improvement, from 32X to 76x speedup, with the new implementation using GPUs over the CPU-based NMFk. Good weak scaling was demonstrated on up to 4096 multi-GPU cluster nodes with approximately 25,000 GPUs when decomposing a dense 340 Terabyte-size matrix and an 11 Exabyte-size sparse matrix of density 10e-6.

arXiv:2202.10435

Survey on Large Scale Neural Network Training

Modern Deep Neural Networks (DNNs) require significant memory to store weight, activations, and other intermediate tensors during training. Hence, many models do not fit one GPU device or can be trained using only a small per-GPU batch size. This survey provides a systematic overview of the approaches that enable more efficient DNNs training. We analyze techniques that save memory and make good use of computation and communication resources on architectures with a single or several GPUs. We summarize the main categories of strategies and compare strategies within and across categories. Along with approaches proposed in the literature, we discuss available implementations.

arXiv:2202.08589

Single UHD Image Dehazing via Interpretable Pyramid Network

Currently, most single image dehazing models cannot run an ultra-high-resolution (UHD) image with a single GPU shader in real-time. To address the problem, we introduce the principle of infinite approximation of Taylor's theorem with the Laplace pyramid pattern to build a model which is capable of handling 4K hazy images in real-time. The N branch networks of the pyramid network correspond to the N constraint terms in Taylor's theorem. Low-order polynomials reconstruct the low-frequency information of the image (e.g. color, illumination). High-order polynomials regress the high-frequency information of the image (e.g. texture). In addition, we propose a Tucker reconstruction-based regularization term that acts on each branch network of the pyramid model. It further constrains the generation of anomalous signals in the feature space. Extensive experimental results demonstrate that our approach can not only run 4K images with haze in real-time on a single GPU (80FPS) but also has unparalleled interpretability. The developed method achieves state-of-the-art (SOTA) performance on two benchmarks (O/I-HAZE) and our updated 4KID dataset while providing the reliable groundwork for subsequent optimization schemes.

arXiv:2202.07896

Aryl: An Elastic Cluster Scheduler for Deep Learning

Companies build separate training and inference GPU clusters for deep learning, and use separate schedulers to manage them. This leads to problems for both training and inference: inference clusters have low GPU utilization when the traffic load is low; training jobs often experience long queueing time due to lack of resources. We introduce Aryl, a new cluster scheduler to address these problems. Aryl introduces capacity loaning to loan idle inference GPU servers for training jobs. It further exploits elastic scaling that scales a training job's GPU allocation to better utilize loaned resources. Capacity loaning and elastic scaling create new challenges to cluster management. When the loaned servers need to be returned, we need to minimize the number of job preemptions; when more GPUs become available, we need to allocate them to elastic jobs and minimize the job completion time (JCT). Aryl addresses these combinatorial problems using principled heuristics. It introduces the notion of server preemption cost which it greedily reduces during server reclaiming. It further relies on the JCT reduction value defined for each additional worker for an elastic job to solve the scheduling problem as a multiple-choice knapsack problem. Prototype implementation on a 64-GPU testbed and large-scale simulation with 15-day traces of over 50,000 production jobs show that Aryl brings 1.53x and 1.50x reductions in average queuing time and JCT, and improves cluster usage by up to 26.9% over the cluster scheduler without capacity loaning or elastic scaling.

arXiv:2202.07798

BB-ML: Basic Block Performance Prediction using Machine Learning Techniques

Recent years have seen the adoption of Machine Learning (ML) techniques to predict the performance of large-scale applications, mostly at a coarse level. In contrast, we propose to use ML techniques for performance prediction at a much finer granularity, namely at the Basic Block (BB) level, which are single entry, single exit code blocks that are used for analysis by the compilers to break down a large code into manageable pieces. We extrapolate the basic block execution counts of GPU applications and use them for predicting the performance for large input sizes from the counts of smaller input sizes. We train a Poisson Neural Network (PNN) model using random input values as well as the lowest input values of the application to learn the relationship between inputs and basic block counts. Experimental results show that the model can accurately predict the basic block execution counts of 16 GPU benchmarks. We achieve an accuracy of 93.5% in extrapolating the basic block counts for large input sets when trained on smaller input sets and an accuracy of 97.7% in predicting basic block counts on random instances. In a case study, we apply the ML model to CUDA GPU benchmarks for performance prediction across a spectrum of applications. We use a variety of metrics for evaluation, including global memory requests and the active cycles of tensor cores, ALU, and FMA units. Results demonstrate the model's capability of predicting the performance of large datasets with an average error rate of 0.85% and 0.17% for global and shared memory requests, respectively. Additionally, to address the utilization of the main functional units in Ampere architecture GPUs, we calculate the active cycles for tensor cores, ALU, FMA, and FP64 units and achieve an average error of 2.3% and 10.66% for ALU and FMA units while the maximum observed error across all tested applications and units reaches 18.5%.

arXiv:2202.07275

HiMA: A Fast and Scalable History-based Memory Access Engine for Differentiable Neural Computer

Memory-augmented neural networks (MANNs) provide better inference performance in many tasks with the help of an external memory. The recently developed differentiable neural computer (DNC) is a MANN that has been shown to outperform in representing complicated data structures and learning long-term dependencies. DNC's higher performance is derived from new history-based attention mechanisms in addition to the previously used content-based attention mechanisms. History-based mechanisms require a variety of new compute primitives and state memories, which are not supported by existing neural network (NN) or MANN accelerators. We present HiMA, a tiled, history-based memory access engine with distributed memories in tiles. HiMA incorporates a multi-mode network-on-chip (NoC) to reduce the communication latency and improve scalability. An optimal submatrix-wise memory partition strategy is applied to reduce the amount of NoC traffic; and a two-stage usage sort method leverages distributed tiles to improve computation speed. To make HiMA fundamentally scalable, we create a distributed version of DNC called DNC-D to allow almost all memory operations to be applied to local memories with trainable weighted summation to produce the global memory output. Two approximation techniques, usage skimming and softmax approximation, are proposed to further enhance hardware efficiency. HiMA prototypes are created in RTL and synthesized in a 40nm technology. By simulations, HiMA running DNC and DNC-D demonstrates 6.47x and 39.1x higher speed, 22.8x and 164.3x better area efficiency, and 6.1x and 61.2x better energy efficiency over the state-of-the-art MANN accelerator. Compared to an Nvidia 3080Ti GPU, HiMA demonstrates speedup by up to 437x and 2,646x when running DNC and DNC-D, respectively.

arXiv:2202.06512공식 코드 · Papers with Code

Benchmarking of DL Libraries and Models on Mobile Devices

Deploying deep learning (DL) on mobile devices has been a notable trend in recent years. To support fast inference of on-device DL, DL libraries play a critical role as algorithms and hardware do. Unfortunately, no prior work ever dives deep into the ecosystem of modern DL libs and provides quantitative results on their performance. In this paper, we first build a comprehensive benchmark that includes 6 representative DL libs and 15 diversified DL models. We then perform extensive experiments on 10 mobile devices, which help reveal a complete landscape of the current mobile DL libs ecosystem. For example, we find that the best-performing DL lib is severely fragmented across different models and hardware, and the gap between those DL libs can be rather huge. In fact, the impacts of DL libs can overwhelm the optimizations from algorithms or hardware, e.g., model quantization and GPU/DSP-based heterogeneous computing. Finally, atop the observations, we summarize practical implications to different roles in the DL lib ecosystem.

arXiv:2202.06819

Learning from distinctive candidates to optimize reduced-precision convolution program on tensor cores

Convolution is one of the fundamental operations of deep neural networks with demanding matrix computation. In a graphic processing unit (GPU), Tensor Core is a specialized matrix processing hardware equipped with reduced-precision matrix-multiply-accumulate (MMA) instructions to increase throughput. However, it is challenging to achieve optimal performance since the best scheduling of MMA instructions varies for different convolution sizes. In particular, reduced-precision MMA requires many elements grouped as a matrix operand, seriously limiting data reuse and imposing packing and layout overhead on the schedule. This work proposes an automatic scheduling method of reduced-precision MMA for convolution operation. In this method, we devise a search space that explores the thread tile and warp sizes to increase the data reuse despite a large matrix operand of reduced-precision MMA. The search space also includes options of register-level packing and layout optimization to lesson overhead of handling reduced-precision data. Finally, we propose a search algorithm to find the best schedule by learning from the distinctive candidates. This reduced-precision MMA optimization method is evaluated on convolution operations of popular neural networks to demonstrate substantial speedup on Tensor Core compared to the state of the arts with shortened search time.

arXiv:2202.02365공식 코드 · Papers with Code

MariusGNN: Resource-Efficient Out-of-Core Training of Graph Neural Networks

We study training of Graph Neural Networks (GNNs) for large-scale graphs. We revisit the premise of using distributed training for billion-scale graphs and show that for graphs that fit in main memory or the SSD of a single machine, out-of-core pipelined training with a single GPU can outperform state-of-the-art (SoTA) multi-GPU solutions. We introduce MariusGNN, the first system that utilizes the entire storage hierarchy -- including disk -- for GNN training. MariusGNN introduces a series of data organization and algorithmic contributions that 1) minimize the end-to-end time required for training and 2) ensure that models learned with disk-based training exhibit accuracy similar to those fully trained in memory. We evaluate MariusGNN against SoTA systems for learning GNN models and find that single-GPU training in MariusGNN achieves the same level of accuracy up to 8x faster than multi-GPU training in these systems, thus, introducing an order of magnitude monetary cost reduction. MariusGNN is open-sourced at www.marius-project.org.

arXiv:2202.03099공식 코드 · Papers with Code

FL_PyTorch: optimization research simulator for federated learning

Federated Learning (FL) has emerged as a promising technique for edge devices to collaboratively learn a shared machine learning model while keeping training data locally on the device, thereby removing the need to store and access the full data in the cloud. However, FL is difficult to implement, test and deploy in practice considering heterogeneity in common edge device settings, making it fundamentally hard for researchers to efficiently prototype and test their optimization algorithms. In this work, our aim is to alleviate this problem by introducing FL_PyTorch : a suite of open-source software written in python that builds on top of one the most popular research Deep Learning (DL) framework PyTorch. We built FL_PyTorch as a research simulator for FL to enable fast development, prototyping and experimenting with new and existing FL optimization algorithms. Our system supports abstractions that provide researchers with a sufficient level of flexibility to experiment with existing and novel approaches to advance the state-of-the-art. Furthermore, FL_PyTorch is a simple to use console system, allows to run several clients simultaneously using local CPUs or GPU(s), and even remote compute devices without the need for any distributed implementation provided by the user. FL_PyTorch also offers a Graphical User Interface. For new methods, researchers only provide the centralized implementation of their algorithm. To showcase the possibilities and usefulness of our system, we experiment with several well-known state-of-the-art FL algorithms and a few of the most common FL datasets.

arXiv:2202.02326공식 코드 · Papers with Code

Towards Training Reproducible Deep Learning Models

Reproducibility is an increasing concern in Artificial Intelligence (AI), particularly in the area of Deep Learning (DL). Being able to reproduce DL models is crucial for AI-based systems, as it is closely tied to various tasks like training, testing, debugging, and auditing. However, DL models are challenging to be reproduced due to issues like randomness in the software (e.g., DL algorithms) and non-determinism in the hardware (e.g., GPU). There are various practices to mitigate some of the aforementioned issues. However, many of them are either too intrusive or can only work for a specific usage context. In this paper, we propose a systematic approach to training reproducible DL models. Our approach includes three main parts: (1) a set of general criteria to thoroughly evaluate the reproducibility of DL models for two different domains, (2) a unified framework which leverages a record-and-replay technique to mitigate software-related randomness and a profile-and-patch technique to control hardware-related non-determinism, and (3) a reproducibility guideline which explains the rationales and the mitigation strategies on conducting a reproducible training process for DL models. Case study results show our approach can successfully reproduce six open source and one commercial DL models.

arXiv:2202.02170공식 코드 · Papers with Code

The Ecological Footprint of Neural Machine Translation Systems

Over the past decade, deep learning (DL) has led to significant advancements in various fields of artificial intelligence, including machine translation (MT). These advancements would not be possible without the ever-growing volumes of data and the hardware that allows large DL models to be trained efficiently. Due to the large amount of computing cores as well as dedicated memory, graphics processing units (GPUs) are a more effective hardware solution for training and inference with DL models than central processing units (CPUs). However, the former is very power demanding. The electrical power consumption has economical as well as ecological implications. This chapter focuses on the ecological footprint of neural MT systems. It starts from the power drain during the training of and the inference with neural MT models and moves towards the environment impact, in terms of carbon dioxide emissions. Different architectures (RNN and Transformer) and different GPUs (consumer-grate NVidia 1080Ti and workstation-grade NVidia P100) are compared. Then, the overall CO2 offload is calculated for Ireland and the Netherlands. The NMT models and their ecological impact are compared to common household appliances to draw a more clear picture. The last part of this chapter analyses quantization, a technique for reducing the size and complexity of models, as a way to reduce power consumption. As quantized models can run on CPUs, they present a power-efficient inference solution without depending on a GPU.

arXiv:2202.01306공식 코드 · Papers with Code

Harmony: Overcoming the Hurdles of GPU Memory Capacity to Train Massive DNN Models on Commodity Servers

Deep neural networks (DNNs) have grown exponentially in size over the past decade, leaving only those who have massive datacenter-based resources with the ability to develop and train such models. One of the main challenges for the long tail of researchers who might have only limited resources (e.g., a single multi-GPU server) is limited GPU memory capacity compared to model size. The problem is so acute that the memory requirement of training massive DNN models can often exceed the aggregate capacity of all available GPUs on a single server; this problem only gets worse with the trend of ever-growing model sizes. Current solutions that rely on virtualizing GPU memory (by swapping to/from CPU memory) incur excessive swapping overhead. In this paper, we present a new training framework, Harmony, and advocate rethinking how DNN frameworks schedule computation and move data to push the boundaries of training massive models efficiently on a single commodity server. Across various massive DNN models, Harmony is able to reduce swap load by up to two orders of magnitude and obtain a training throughput speedup of up to 7.6x over highly optimized baselines with virtualized memory.

arXiv:2202.01258공식 코드 · Papers with Code

Accelerated Quality-Diversity through Massive Parallelism

Quality-Diversity (QD) optimization algorithms are a well-known approach to generate large collections of diverse and high-quality solutions. However, derived from evolutionary computation, QD algorithms are population-based methods which are known to be data-inefficient and requires large amounts of computational resources. This makes QD algorithms slow when used in applications where solution evaluations are computationally costly. A common approach to speed up QD algorithms is to evaluate solutions in parallel, for instance by using physical simulators in robotics. Yet, this approach is limited to several dozen of parallel evaluations as most physics simulators can only be parallelized more with a greater number of CPUs. With recent advances in simulators that run on accelerators, thousands of evaluations can now be performed in parallel on single GPU/TPU. In this paper, we present QDax, an accelerated implementation of MAP-Elites which leverages massive parallelism on accelerators to make QD algorithms more accessible. We show that QD algorithms are ideal candidates to take advantage of progress in hardware acceleration. We demonstrate that QD algorithms can scale with massive parallelism to be run at interactive timescales without any significant effect on the performance. Results across standard optimization functions and four neuroevolution benchmark environments shows that experiment runtimes are reduced by two factors of magnitudes, turning days of computation into minutes. More surprising, we observe that reducing the number of generations by two orders of magnitude, and thus having significantly shorter lineage does not impact the performance of QD algorithms. These results show that QD can now benefit from hardware acceleration, which contributed significantly to the bloom of deep learning.

arXiv:2202.01085공식 코드 · Papers with Code

Giga-scale Kernel Matrix Vector Multiplication on GPU

Kernel matrix-vector multiplication (KMVM) is a foundational operation in machine learning and scientific computing. However, as KMVM tends to scale quadratically in both memory and time, applications are often limited by these computational constraints. In this paper, we propose a novel approximation procedure coined \textit{Faster-Fast and Free Memory Method} ($\fthreem$) to address these scaling issues of KMVM for tall~($10^8\sim 10^9$) and skinny~($D\leq7$) data. Extensive experiments demonstrate that $\fthreem$ has empirical \emph{linear time and memory} complexity with a relative error of order $10^{-3}$ and can compute a full KMVM for a billion points \emph{in under a minute} on a high-end GPU, leading to a significant speed-up in comparison to existing CPU methods. We demonstrate the utility of our procedure by applying it as a drop-in for the state-of-the-art GPU-based linear solver FALKON, \emph{improving speed 1.5-5.5 times} at the cost of $<1\%$ drop in accuracy. We further demonstrate competitive results on \emph{Gaussian Process regression} coupled with significant speedups on a variety of real-world datasets.