PyTorchKR Papers

최신 논문

총 5,576편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2202.00774공식 코드 · Papers with Code

Accelerating DNN Training with Structured Data Gradient Pruning

Weight pruning is a technique to make Deep Neural Network (DNN) inference more computationally efficient by reducing the number of model parameters over the course of training. However, most weight pruning techniques generally does not speed up DNN training and can even require more iterations to reach model convergence. In this work, we propose a novel Structured Data Gradient Pruning (SDGP) method that can speed up training without impacting model convergence. This approach enforces a specific sparsity structure, where only N out of every M elements in a matrix can be nonzero, making it amenable to hardware acceleration. Modern accelerators such as the Nvidia A100 GPU support this type of structured sparsity for 2 nonzeros per 4 elements in a reduction. Assuming hardware support for 2:4 sparsity, our approach can achieve a 15-25\% reduction in total training time without significant impact to performance. Source code and pre-trained models are available at \url{https://github.com/BradMcDanel/sdgp}.

arXiv:2202.00874공식 코드 · Papers with Code

HTS-AT: A Hierarchical Token-Semantic Audio Transformer for Sound Classification and Detection

Audio classification is an important task of mapping audio samples into their corresponding labels. Recently, the transformer model with self-attention mechanisms has been adopted in this field. However, existing audio transformers require large GPU memories and long training time, meanwhile relying on pretrained vision models to achieve high performance, which limits the model's scalability in audio tasks. To combat these problems, we introduce HTS-AT: an audio transformer with a hierarchical structure to reduce the model size and training time. It is further combined with a token-semantic module to map final outputs into class featuremaps, thus enabling the model for the audio event detection (i.e. localization in time). We evaluate HTS-AT on three datasets of audio classification where it achieves new state-of-the-art (SOTA) results on AudioSet and ESC-50, and equals the SOTA on Speech Command V2. It also achieves better performance in event localization than the previous CNN-based models. Moreover, HTS-AT requires only 35% model parameters and 15% training time of the previous audio transformer. These results demonstrate the high performance and high efficiency of HTS-AT.

arXiv:2201.13309

Accelerating Laue Depth Reconstruction Algorithm with CUDA

The Laue diffraction microscopy experiment uses the polychromatic Laue micro-diffraction technique to examine the structure of materials with sub-micron spatial resolution in all three dimensions. During this experiment, local crystallographic orientations, orientation gradients and strains are measured as properties which will be recorded in HDF5 image format. The recorded images will be processed with a depth reconstruction algorithm for future data analysis. But the current depth reconstruction algorithm consumes considerable processing time and might take up to 2 weeks for reconstructing data collected from one single experiment. To improve the depth reconstruction computation speed, we propose a scalable GPU program solution on the depth reconstruction problem in this paper. The test result shows that the running time would be 10 to 20 times faster than the prior CPU design for various size of input data.

arXiv:2201.13191

Computational Scatter Correction for High-Resolution Flat-Panel CT Based on a Fast Monte Carlo Photon Transport Model

In computed tomography (CT) reconstruction, scattering causes server quality degradation of the reconstructed CT images by introducing streaks and cupping artifacts which reduce the detectability of low contrast objects. Monte Carlo (MC) simulation is considered as the most accurate approach for scatter estimation. However, the existing MC estimators are computationally expensive especially for the considered high-resolution flat-panel CT. In this paper, we propose a fast and accurate photon transport model which describes the physics within the 1 keV to 1 MeV range using multiple controllable key parameters. Based on this model, scatter computation for a single projection can be completed within a range of few seconds under well-defined model parameters. Smoothing and interpolation are performed on the estimated scatter to accelerate the scatter calculation without compromising accuracy too much compared to measured near scatter-free projection images. Combining the scatter estimation with the filtered backprojection (FBP), scatter correction is performed effectively in an iterative manner. In order to evaluate the proposed MC model, we have conducted extensive experiments on the simulated data and real-world high-resolution flat-panel CT. Comparing to the state-of-the-art MC simulators, our photon transport model achieved a 202$\times$ speed-up on a four GPU system comparing to the multi-threaded state-of-the-art EGSnrc MC simulator. Besides, it is shown that for real-world high-resolution flat-panel CT, scatter correction with sufficient accuracy is accomplished within one to three iterations using a FBP and a forward projection computed with the proposed fast MC photon transport model.

arXiv:2201.12423

Benchmarking Resource Usage for Efficient Distributed Deep Learning

Deep learning (DL) workflows demand an ever-increasing budget of compute and energy in order to achieve outsized gains. Neural architecture searches, hyperparameter sweeps, and rapid prototyping consume immense resources that can prevent resource-constrained researchers from experimenting with large models and carry considerable environmental impact. As such, it becomes essential to understand how different deep neural networks (DNNs) and training leverage increasing compute and energy resources -- especially specialized computationally-intensive models across different domains and applications. In this paper, we conduct over 3,400 experiments training an array of deep networks representing various domains/tasks -- natural language processing, computer vision, and chemistry -- on up to 424 graphics processing units (GPUs). During training, our experiments systematically vary compute resource characteristics and energy-saving mechanisms such as power utilization and GPU clock rate limits to capture and illustrate the different trade-offs and scaling behaviors each representative model exhibits under various resource and energy-constrained regimes. We fit power law models that describe how training time scales with available compute resources and energy constraints. We anticipate that these findings will help inform and guide high-performance computing providers in optimizing resource utilization, by selectively reducing energy consumption for different deep learning tasks/workflows with minimal impact on training.

arXiv:2201.13096공식 코드 · Papers with Code

SPDY: Accurate Pruning with Speedup Guarantees

The recent focus on the efficiency of deep neural networks (DNNs) has led to significant work on model compression approaches, of which weight pruning is one of the most popular. At the same time, there is rapidly-growing computational support for efficiently executing the unstructured-sparse models obtained via pruning. Yet, most existing pruning methods minimize just the number of remaining weights, i.e. the size of the model, rather than optimizing for inference time. We address this gap by introducing SPDY, a new compression method which automatically determines layer-wise sparsity targets achieving a desired inference speedup on a given system, while minimizing accuracy loss. SPDY is composed of two new techniques: the first is an efficient dynamic programming algorithm for solving the speedup-constrained layer-wise compression problem assuming a set of given layer-wise sensitivity scores; the second is a local search procedure for determining accurate layer-wise sensitivity scores. Experiments across popular vision and language models show that SPDY guarantees speedups while recovering higher accuracy relative to existing strategies, both for one-shot and gradual pruning scenarios, and is compatible with most existing pruning approaches. We also extend our approach to the recently-proposed task of pruning with very little data, where we achieve the best known accuracy recovery when pruning to the GPU-supported 2:4 sparsity pattern.

arXiv:2201.13278

Combining Local and Global Pose Estimation for Precise Tracking of Similar Objects

In this paper, we present a multi-object 6D detection and tracking pipeline for potentially similar and non-textured objects. The combination of a convolutional neural network for object classification and rough pose estimation with a local pose refinement and an automatic mismatch detection enables direct application in real-time AR scenarios. A new network architecture, trained solely with synthetic images, allows simultaneous pose estimation of multiple objects with reduced GPU memory consumption and enhanced performance. In addition, the pose estimates are further improved by a local edge-based refinement step that explicitly exploits known object geometry information. For continuous movements, the sole use of local refinement reduces pose mismatches due to geometric ambiguities or occlusions. We showcase the entire tracking pipeline and demonstrate the benefits of the combined approach. Experiments on a challenging set of non-textured similar objects demonstrate the enhanced quality compared to the baseline method. Finally, we illustrate how the system can be used in a real AR assistance application within the field of construction.

arXiv:2112.11925

SOLIS -- The MLOps journey from data acquisition to actionable insights

Machine Learning operations is unarguably a very important and also one of the hottest topics in Artificial Intelligence lately. Being able to define very clear hypotheses for actual real-life problems that can be addressed by machine learning models, collecting and curating large amounts of data for model training and validation followed by model architecture search and actual optimization and finally presenting the results fits very well the scenario of Data Science experiments. This approach however does not supply the needed procedures and pipelines for the actual deployment of machine learning capabilities in real production grade systems. Automating live configuration mechanisms, on the fly adapting to live or offline data capture and consumption, serving multiple models in parallel either on edge or cloud architectures, addressing specific limitations of GPU memory or compute power, post-processing inference or prediction results and serving those either as APIs or with IoT based communication stacks in the same end-to-end pipeline are the real challenges that we try to address in this particular paper. In this paper we present a unified deployment pipeline and freedom-to-operate approach that supports all above requirements while using basic cross-platform tensor framework and script language engines.

arXiv:2201.11853

Prediction of GPU Failures Under Deep Learning Workloads

Graphics processing units (GPUs) are the de facto standard for processing deep learning (DL) tasks. Meanwhile, GPU failures, which are inevitable, cause severe consequences in DL tasks: they disrupt distributed trainings, crash inference services, and result in service level agreement violations. To mitigate the problem caused by GPU failures, we propose to predict failures by using ML models. This paper is the first to study prediction models of GPU failures under large-scale production deep learning workloads. As a starting point, we evaluate classic prediction models and observe that predictions of these models are both inaccurate and unstable. To improve the precision and stability of predictions, we propose several techniques, including parallel and cascade model-ensemble mechanisms and a sliding training method. We evaluate the performances of our various techniques on a four-month production dataset including 350 million entries. The results show that our proposed techniques improve the prediction precision from 46.3\% to 84.0\%.

arXiv:2201.10781

ASFD: Automatic and Scalable Face Detector

Along with current multi-scale based detectors, Feature Aggregation and Enhancement (FAE) modules have shown superior performance gains for cutting-edge object detection. However, these hand-crafted FAE modules show inconsistent improvements on face detection, which is mainly due to the significant distribution difference between its training and applying corpus, COCO vs. WIDER Face. To tackle this problem, we essentially analyse the effect of data distribution, and consequently propose to search an effective FAE architecture, termed AutoFAE by a differentiable architecture search, which outperforms all existing FAE modules in face detection with a considerable margin. Upon the found AutoFAE and existing backbones, a supernet is further built and trained, which automatically obtains a family of detectors under the different complexity constraints. Extensive experiments conducted on popular benchmarks, WIDER Face and FDDB, demonstrate the state-of-the-art performance-efficiency trade-off for the proposed automatic and scalable face detector (ASFD) family. In particular, our strong ASFD-D6 outperforms the best competitor with AP 96.7/96.2/92.1 on WIDER Face test, and the lightweight ASFD-D0 costs about 3.1 ms, more than 320 FPS, on the V100 GPU with VGA-resolution images.

arXiv:2201.10271공식 코드 · Papers with Code

Convolutional Xformers for Vision

Vision transformers (ViTs) have found only limited practical use in processing images, in spite of their state-of-the-art accuracy on certain benchmarks. The reason for their limited use include their need for larger training datasets and more computational resources compared to convolutional neural networks (CNNs), owing to the quadratic complexity of their self-attention mechanism. We propose a linear attention-convolution hybrid architecture -- Convolutional X-formers for Vision (CXV) -- to overcome these limitations. We replace the quadratic attention with linear attention mechanisms, such as Performer, Nystr\"omformer, and Linear Transformer, to reduce its GPU usage. Inductive prior for image data is provided by convolutional sub-layers, thereby eliminating the need for class token and positional embeddings used by the ViTs. We also propose a new training method where we use two different optimizers during different phases of training and show that it improves the top-1 image classification accuracy across different architectures. CXV outperforms other architectures, token mixers (e.g. ConvMixer, FNet and MLP Mixer), transformer models (e.g. ViT, CCT, CvT and hybrid Xformers), and ResNets for image classification in scenarios with limited data and GPU resources (cores, RAM, power).

arXiv:2201.09377

An Application of Pseudo-Log-Likelihoods to Natural Language Scoring

Language models built using semi-supervised machine learning on large corpora of natural language have very quickly enveloped the fields of natural language generation and understanding. In this paper we apply a zero-shot approach independently developed by a number of researchers now gaining recognition as a significant alternative to fine-tuning for evaluation on common sense tasks. A language model with relatively few parameters and training steps compared to a more recent language model (T5) can outperform it on a recent large data set (TimeDial), while displaying robustness in its performance across a similar class of language tasks. Surprisingly, this result is achieved by using a hyperparameter-free zero-shot method with the smaller model, compared to fine-tuning to the larger model. We argue that robustness of the smaller model ought to be understood in terms of compositionality, in a sense that we draw from recent literature on a class of similar models. We identify a practical cost for our method and model: high GPU-time for natural language evaluation. The zero-shot measurement technique that produces remarkable stability, both for ALBERT and other BERT variants, is an application of pseudo-log-likelihoods to masked language models for the relative measurement of probability for substitution alternatives in forced choice language tasks such as the Winograd Schema Challenge, Winogrande, and others. One contribution of this paper is to bring together a number of similar, but independent strands of research. We produce some absolute state-of-the-art results for common sense reasoning in binary choice tasks, performing better than any published result in the literature, including fine-tuned efforts. We show a remarkable consistency of the model's performance under adversarial settings, which we argue is best explained by the model's compositionality of representations.

arXiv:2201.09240

Learning-Driven Lossy Image Compression; A Comprehensive Survey

In the realm of image processing and computer vision (CV), machine learning (ML) architectures are widely applied. Convolutional neural networks (CNNs) solve a wide range of image processing issues and can solve image compression problem. Compression of images is necessary due to bandwidth and memory constraints. Helpful, redundant, and irrelevant information are three different forms of information found in images. This paper aims to survey recent techniques utilizing mostly lossy image compression using ML architectures including different auto-encoders (AEs) such as convolutional auto-encoders (CAEs), variational auto-encoders (VAEs), and AEs with hyper-prior models, recurrent neural networks (RNNs), CNNs, generative adversarial networks (GANs), principal component analysis (PCA) and fuzzy means clustering. We divide all of the algorithms into several groups based on architecture. We cover still image compression in this survey. Various discoveries for the researchers are emphasized and possible future directions for researchers. The open research problems such as out of memory (OOM), striped region distortion (SRD), aliasing, and compatibility of the frameworks with central processing unit (CPU) and graphics processing unit (GPU) simultaneously are explained. The majority of the publications in the compression domain surveyed are from the previous five years and use a variety of approaches.

arXiv:2201.08475공식 코드 · Papers with Code

GenGNN: A Generic FPGA Framework for Graph Neural Network Acceleration

Graph neural networks (GNNs) have recently exploded in popularity thanks to their broad applicability to ubiquitous graph-related problems such as quantum chemistry, drug discovery, and high energy physics. However, meeting demand for novel GNN models and fast inference simultaneously is challenging because of the gap between the difficulty in developing efficient FPGA accelerators and the rapid pace of creation of new GNN models. Prior art focuses on the acceleration of specific classes of GNNs but lacks the generality to work across existing models or to extend to new and emerging GNN models. In this work, we propose a generic GNN acceleration framework using High-Level Synthesis (HLS), named GenGNN, with two-fold goals. First, we aim to deliver ultra-fast GNN inference without any graph pre-processing for real-time requirements. Second, we aim to support a diverse set of GNN models with the extensibility to flexibly adapt to new models. The framework features an optimized message-passing structure applicable to all models, combined with a rich library of model-specific components. We verify our implementation on-board on the Xilinx Alveo U50 FPGA and observe a speed-up of up to 25x against CPU (6226R) baseline and 13x against GPU (A6000) baseline. Our HLS code will be open-source on GitHub upon acceptance.

arXiv:2201.07821공식 코드 · Papers with Code

Building a Performance Model for Deep Learning Recommendation Model Training on GPUs

We devise a performance model for GPU training of Deep Learning Recommendation Models (DLRM), whose GPU utilization is low compared to other well-optimized CV and NLP models. We show that both the device active time (the sum of kernel runtimes) but also the device idle time are important components of the overall device time. We therefore tackle them separately by (1) flexibly adopting heuristic-based and ML-based kernel performance models for operators that dominate the device active time, and (2) categorizing operator overheads into five types to determine quantitatively their contribution to the device active time. Combining these two parts, we propose a critical-path-based algorithm to predict the per-batch training time of DLRM by traversing its execution graph. We achieve less than 10% geometric mean average error (GMAE) in all kernel performance modeling, and 4.61% and 7.96% geomean errors for GPU active time and overall E2E per-batch training time prediction with overheads from individual workloads, respectively. A slight increase of 2.19% incurred in E2E prediction error with shared overheads across workloads suggests the feasibility of using shared overheads in large-scale prediction. We show that our general performance model not only achieves low prediction error on DLRM, which has highly customized configurations and is dominated by multiple factors but also yields comparable accuracy on other compute-bound ML models targeted by most previous methods. Using this performance model and graph-level data and task dependency analysis, we show our system can provide more general model-system co-design than previous methods.

arXiv:2201.08098

What can we learn from misclassified ImageNet images?

Understanding the patterns of misclassified ImageNet images is particularly important, as it could guide us to design deep neural networks (DNN) that generalize better. However, the richness of ImageNet imposes difficulties for researchers to visually find any useful patterns of misclassification. Here, to help find these patterns, we propose "Superclassing ImageNet dataset". It is a subset of ImageNet which consists of 10 superclasses, each containing 7-116 related subclasses (e.g., 52 bird types, 116 dog types). By training neural networks on this dataset, we found that: (i) Misclassifications are rarely across superclasses, but mainly among subclasses within a superclass. (ii) Ensemble networks trained each only on subclasses of a given superclass perform better than the same network trained on all subclasses of all superclasses. Hence, we propose a two-stage Super-Sub framework, and demonstrate that: (i) The framework improves overall classification performance by 3.3%, by first inferring a superclass using a generalist superclass-level network, and then using a specialized network for final subclass-level classification. (ii) Although the total parameter storage cost increases to a factor N+1 for N superclasses compared to using a single network, with finetuning, delta and quantization aware training techniques this can be reduced to 0.2N+1. Another advantage of this efficient implementation is that the memory cost on the GPU during inference is equivalent to using only one network. The reason is we initiate each subclass-level network through addition of small parameter variations (deltas) to the superclass-level network. (iii) Finally, our framework promises to be more scalable and generalizable than the common alternative of simply scaling up a vanilla network in size, since very large networks often suffer from overfitting and gradient vanishing.

arXiv:2201.07562

Learned Cone-Beam CT Reconstruction Using Neural Ordinary Differential Equations

Learned iterative reconstruction algorithms for inverse problems offer the flexibility to combine analytical knowledge about the problem with modules learned from data. This way, they achieve high reconstruction performance while ensuring consistency with the measured data. In computed tomography, extending such approaches from 2D fan-beam to 3D cone-beam data is challenging due to the prohibitively high GPU memory that would be needed to train such models. This paper proposes to use neural ordinary differential equations to solve the reconstruction problem in a residual formulation via numerical integration. For training, there is no need to backpropagate through several unrolled network blocks nor through the internals of the solver. Instead, the gradients are obtained very memory-efficiently in the neural ODE setting allowing for training on a single consumer graphics card. The method is able to reduce the root mean squared error by over 30% compared to the best performing classical iterative reconstruction algorithm and produces high quality cone-beam reconstructions even in a sparse view scenario.

arXiv:2201.07210

Efficient Training of Spiking Neural Networks with Temporally-Truncated Local Backpropagation through Time

Directly training spiking neural networks (SNNs) has remained challenging due to complex neural dynamics and intrinsic non-differentiability in firing functions. The well-known backpropagation through time (BPTT) algorithm proposed to train SNNs suffers from large memory footprint and prohibits backward and update unlocking, making it impossible to exploit the potential of locally-supervised training methods. This work proposes an efficient and direct training algorithm for SNNs that integrates a locally-supervised training method with a temporally-truncated BPTT algorithm. The proposed algorithm explores both temporal and spatial locality in BPTT and contributes to significant reduction in computational cost including GPU memory utilization, main memory access and arithmetic operations. We thoroughly explore the design space concerning temporal truncation length and local training block size and benchmark their impact on classification accuracy of different networks running different types of tasks. The results reveal that temporal truncation has a negative effect on the accuracy of classifying frame-based datasets, but leads to improvement in accuracy on dynamic-vision-sensor (DVS) recorded datasets. In spite of resulting information loss, local training is capable of alleviating overfitting. The combined effect of temporal truncation and local training can lead to the slowdown of accuracy drop and even improvement in accuracy. In addition, training deep SNNs models such as AlexNet classifying CIFAR10-DVS dataset leads to 7.26% increase in accuracy, 89.94% reduction in GPU memory, 10.79% reduction in memory access, and 99.64% reduction in MAC operations compared to the standard end-to-end BPTT.

arXiv:2201.07705

GEMEL: Model Merging for Memory-Efficient, Real-Time Video Analytics at the Edge

Video analytics pipelines have steadily shifted to edge deployments to reduce bandwidth overheads and privacy violations, but in doing so, face an ever-growing resource tension. Most notably, edge-box GPUs lack the memory needed to concurrently house the growing number of (increasingly complex) models for real-time inference. Unfortunately, existing solutions that rely on time/space sharing of GPU resources are insufficient as the required swapping delays result in unacceptable frame drops and accuracy violations. We present model merging, a new memory management technique that exploits architectural similarities between edge vision models by judiciously sharing their layers (including weights) to reduce workload memory costs and swapping delays. Our system, GEMEL, efficiently integrates merging into existing pipelines by (1) leveraging several guiding observations about per-model memory usage and inter-layer dependencies to quickly identify fruitful and accuracy-preserving merging configurations, and (2) altering edge inference schedules to maximize merging benefits. Experiments across diverse workloads reveal that GEMEL reduces memory usage by up to 60.7%, and improves overall accuracy by 8-39% relative to time/space sharing alone.

arXiv:2201.07692

GroupGazer: A Tool to Compute the Gaze per Participant in Groups with integrated Calibration to Map the Gaze Online to a Screen or Beamer Projection

In this paper we present GroupGaze. It is a tool that can be used to calculate the gaze direction and the gaze position of whole groups. GroupGazer calculates the gaze direction of every single person in the image and allows to map these gaze vectors to a projection like a projector. In addition to the person-specific gaze direction, the person affiliation of each gaze vector is stored based on the position in the image. Also, it is possible to save the group attention after a calibration. The software is free to use and requires a simple webcam as well as an NVIDIA GPU and the operating system Windows or Linux. Link: https://es-cloud.cs.uni-tuebingen.de/d/8e2ab8c3fdd444e1a135/?p=%2FGroupGazer&mode=list