PyTorchKR Papers

최신 논문

총 5,576편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2106.06126

Exploiting Large-scale Teacher-Student Training for On-device Acoustic Models

We present results from Alexa speech teams on semi-supervised learning (SSL) of acoustic models (AM) with experiments spanning over 3000 hours of GPU time, making our study one of the largest of its kind. We discuss SSL for AMs in a small footprint setting, showing that a smaller capacity model trained with 1 million hours of unsupervised data can outperform a baseline supervised system by 14.3% word error rate reduction (WERR). When increasing the supervised data to seven-fold, our gains diminish to 7.1% WERR; to improve SSL efficiency at larger supervised data regimes, we employ a step-wise distillation into a smaller model, obtaining a WERR of 14.4%. We then switch to SSL using larger student models in low data regimes; while learning efficiency with unsupervised data is higher, student models may outperform teacher models in such a setting. We develop a theoretical sketch to explain this behavior.

arXiv:2106.05825

HASI: Hardware-Accelerated Stochastic Inference, A Defense Against Adversarial Machine Learning Attacks

Deep Neural Networks (DNNs) are employed in an increasing number of applications, some of which are safety critical. Unfortunately, DNNs are known to be vulnerable to so-called adversarial attacks that manipulate inputs to cause incorrect results that can be beneficial to an attacker or damaging to the victim. Multiple defenses have been proposed to increase the robustness of DNNs. In general, these defenses have high overhead, some require attack-specific re-training of the model or careful tuning to adapt to different attacks. This paper presents HASI, a hardware-accelerated defense that uses a process we call stochastic inference to detect adversarial inputs. We show that by carefully injecting noise into the model at inference time, we can differentiate adversarial inputs from benign ones. HASI uses the output distribution characteristics of noisy inference compared to a non-noisy reference to detect adversarial inputs. We show an adversarial detection rate of 86% when applied to VGG16 and 93% when applied to ResNet50, which exceeds the detection rate of the state of the art approaches, with a much lower overhead. We demonstrate two software/hardware-accelerated co-designs, which reduces the performance impact of stochastic inference to 1.58X-2X relative to the unprotected baseline, compared to 15X-20X overhead for a software-only GPU implementation.

arXiv:2106.05609공식 코드 · Papers with Code

GNNAutoScale: Scalable and Expressive Graph Neural Networks via Historical Embeddings

We present GNNAutoScale (GAS), a framework for scaling arbitrary message-passing GNNs to large graphs. GAS prunes entire sub-trees of the computation graph by utilizing historical embeddings from prior training iterations, leading to constant GPU memory consumption in respect to input node size without dropping any data. While existing solutions weaken the expressive power of message passing due to sub-sampling of edges or non-trainable propagations, our approach is provably able to maintain the expressive power of the original GNN. We achieve this by providing approximation error bounds of historical embeddings and show how to tighten them in practice. Empirically, we show that the practical realization of our framework, PyGAS, an easy-to-use extension for PyTorch Geometric, is both fast and memory-efficient, learns expressive node representations, closely resembles the performance of their non-scaling counterparts, and reaches state-of-the-art performance on large-scale graphs.

arXiv:2106.05611

Context-Free TextSpotter for Real-Time and Mobile End-to-End Text Detection and Recognition

In the deployment of scene-text spotting systems on mobile platforms, lightweight models with low computation are preferable. In concept, end-to-end (E2E) text spotting is suitable for such purposes because it performs text detection and recognition in a single model. However, current state-of-the-art E2E methods rely on heavy feature extractors, recurrent sequence modellings, and complex shape aligners to pursue accuracy, which means their computations are still heavy. We explore the opposite direction: How far can we go without bells and whistles in E2E text spotting? To this end, we propose a text-spotting method that consists of simple convolutions and a few post-processes, named Context-Free TextSpotter. Experiments using standard benchmarks show that Context-Free TextSpotter achieves real-time text spotting on a GPU with only three million parameters, which is the smallest and fastest among existing deep text spotters, with an acceptable transcription quality degradation compared to heavier ones. Further, we demonstrate that our text spotter can run on a smartphone with affordable latency, which is valuable for building stand-alone OCR applications.

arXiv:2106.04784공식 코드 · Papers with Code

Accelerating Neural Architecture Search via Proxy Data

Despite the increasing interest in neural architecture search (NAS), the significant computational cost of NAS is a hindrance to researchers. Hence, we propose to reduce the cost of NAS using proxy data, i.e., a representative subset of the target data, without sacrificing search performance. Even though data selection has been used across various fields, our evaluation of existing selection methods for NAS algorithms offered by NAS-Bench-1shot1 reveals that they are not always appropriate for NAS and a new selection method is necessary. By analyzing proxy data constructed using various selection methods through data entropy, we propose a novel proxy data selection method tailored for NAS. To empirically demonstrate the effectiveness, we conduct thorough experiments across diverse datasets, search spaces, and NAS algorithms. Consequently, NAS algorithms with the proposed selection discover architectures that are competitive with those obtained using the entire dataset. It significantly reduces the search cost: executing DARTS with the proposed selection requires only 40 minutes on CIFAR-10 and 7.5 hours on ImageNet with a single GPU. Additionally, when the architecture searched on ImageNet using the proposed selection is inversely transferred to CIFAR-10, a state-of-the-art test error of 2.4\% is yielded. Our code is available at https://github.com/nabk89/NAS-with-Proxy-data.

논문 ID: document/8593800

SOS: Stereo Matching in O(1) with Slanted Support Windows

Depth cameras have accelerated research in many areas of computer vision. Most triangulation-based depth cameras, whether structured light systems like the Kinect or active (assisted) stereo systems, are based on the principle of stereo matching. Depth from stereo is an active research topic dating back 30 years. Despite recent advances, algorithms usually trade-off accuracy for speed. In particular, efficient methods rely on fronto-parallel assumptions to reduce the search space and keep computation low. We present SOS (Slanted O(1) Stereo), the first algorithm capable of leveraging slanted support windows without sacrificing speed or accuracy. We use an active stereo configuration, where an illuminator textures the scene. Under this setting, local methods - such as PatchMatch Stereo - obtain state of the art results by jointly estimating disparities and slant, but at a large computational cost. We observe that these methods typically exploit local smoothness to simplify their initialization strategies. Our key insight is that local smoothness can in fact be used to amortize the computation not only within initialization, but across the entire stereo pipeline. Building on these insights, we propose a novel hierarchical initialization that is able to efficiently perform search over disparity and slants. We then show how this structure can be leveraged to provide high quality depth maps. Extensive quantitative evaluations demonstrate that the proposed technique yields significantly more precise results than current state of the art, but at a fraction of the computational cost. Our prototype implementation runs at 4000 fps on modern GPU architectures.

arXiv:2106.03167

Mathematical Vocoder Algorithm : Modified Spectral Inversion for Efficient Neural Speech Synthesis

In this work, we propose a new mathematical vocoder algorithm(modified spectral inversion) that generates a waveform from acoustic features without phase estimation. The main benefit of using our proposed method is that it excludes the training stage of the neural vocoder from the end-to-end speech synthesis model. Our implementation can synthesize high fidelity speech at approximately 20 Mhz on CPU and 59.6MHz on GPU. This is 909 and 2,702 times faster compared to real-time. Since the proposed methodology is not a data-driven method, it is applicable to unseen voices and multiple languages without any additional work. The proposed method is expected to adapt for researching on neural network models capable of synthesizing speech at the studio recording level.

arXiv:2106.03287공식 코드 · Papers with Code

Stein ICP for Uncertainty Estimation in Point Cloud Matching

Quantification of uncertainty in point cloud matching is critical in many tasks such as pose estimation, sensor fusion, and grasping. Iterative closest point (ICP) is a commonly used pose estimation algorithm which provides a point estimate of the transformation between two point clouds. There are many sources of uncertainty in this process that may arise due to sensor noise, ambiguous environment, and occlusion. However, for safety critical problems such as autonomous driving, a point estimate of the pose transformation is not sufficient as it does not provide information about the multiple solutions. Current probabilistic ICP methods usually do not capture all sources of uncertainty and may provide unreliable transformation estimates which can have a detrimental effect in state estimation or decision making tasks that use this information. In this work we propose a new algorithm to align two point clouds that can precisely estimate the uncertainty of ICP's transformation parameters. We develop a Stein variational inference framework with gradient based optimization of ICP's cost function. The method provides a non-parametric estimate of the transformation, can model complex multi-modal distributions, and can be effectively parallelized on a GPU. Experiments using 3D kinect data as well as sparse indoor/outdoor LiDAR data show that our method is capable of efficiently producing accurate pose uncertainty estimates.

arXiv:2106.02694공식 코드 · Papers with Code

Efficient Classification of Very Large Images with Tiny Objects

An increasing number of applications in computer vision, specially, in medical imaging and remote sensing, become challenging when the goal is to classify very large images with tiny informative objects. Specifically, these classification tasks face two key challenges: $i$) the size of the input image is usually in the order of mega- or giga-pixels, however, existing deep architectures do not easily operate on such big images due to memory constraints, consequently, we seek a memory-efficient method to process these images; and $ii$) only a very small fraction of the input images are informative of the label of interest, resulting in low region of interest (ROI) to image ratio. However, most of the current convolutional neural networks (CNNs) are designed for image classification datasets that have relatively large ROIs and small image sizes (sub-megapixel). Existing approaches have addressed these two challenges in isolation. We present an end-to-end CNN model termed Zoom-In network that leverages hierarchical attention sampling for classification of large images with tiny objects using a single GPU. We evaluate our method on four large-image histopathology, road-scene and satellite imaging datasets, and one gigapixel pathology dataset. Experimental results show that our model achieves higher accuracy than existing methods while requiring less memory resources.

arXiv:2106.03112공식 코드 · Papers with Code

Rethinking Training from Scratch for Object Detection

The ImageNet pre-training initialization is the de-facto standard for object detection. He et al. found it is possible to train detector from scratch(random initialization) while needing a longer training schedule with proper normalization technique. In this paper, we explore to directly pre-training on target dataset for object detection. Under this situation, we discover that the widely adopted large resizing strategy e.g. resize image to (1333, 800) is important for fine-tuning but it's not necessary for pre-training. Specifically, we propose a new training pipeline for object detection that follows `pre-training and fine-tuning', utilizing low resolution images within target dataset to pre-training detector then load it to fine-tuning with high resolution images. With this strategy, we can use batch normalization(BN) with large bath size during pre-training, it's also memory efficient that we can apply it on machine with very limited GPU memory(11G). We call it direct detection pre-training, and also use direct pre-training for short. Experiment results show that direct pre-training accelerates the pre-training phase by more than 11x on COCO dataset while with even +1.8mAP compared to ImageNet pre-training. Besides, we found direct pre-training is also applicable to transformer based backbones e.g. Swin Transformer. Code will be available.

arXiv:2106.03452공식 코드 · Papers with Code

Shape As Points: A Differentiable Poisson Solver

In recent years, neural implicit representations gained popularity in 3D reconstruction due to their expressiveness and flexibility. However, the implicit nature of neural implicit representations results in slow inference time and requires careful initialization. In this paper, we revisit the classic yet ubiquitous point cloud representation and introduce a differentiable point-to-mesh layer using a differentiable formulation of Poisson Surface Reconstruction (PSR) that allows for a GPU-accelerated fast solution of the indicator function given an oriented point cloud. The differentiable PSR layer allows us to efficiently and differentiably bridge the explicit 3D point representation with the 3D mesh via the implicit indicator field, enabling end-to-end optimization of surface reconstruction metrics such as Chamfer distance. This duality between points and meshes hence allows us to represent shapes as oriented point clouds, which are explicit, lightweight and expressive. Compared to neural implicit representations, our Shape-As-Points (SAP) model is more interpretable, lightweight, and accelerates inference time by one order of magnitude. Compared to other explicit representations such as points, patches, and meshes, SAP produces topology-agnostic, watertight manifold surfaces. We demonstrate the effectiveness of SAP on the task of surface reconstruction from unoriented point clouds and learning-based reconstruction.

arXiv:2106.02679

Layered gradient accumulation and modular pipeline parallelism: fast and efficient training of large language models

The advent of the transformer has sparked a quick growth in the size of language models, far outpacing hardware improvements. (Dense) transformers are expected to reach the trillion-parameter scale in the near future, for which training requires thousands or even tens of thousands of GPUs. We investigate the challenges of training at this scale and beyond on commercially available hardware. In particular, we analyse the shortest possible training time for different configurations of distributed training, leveraging empirical scaling laws for language models to estimate the optimal (critical) batch size. Contrary to popular belief, we find no evidence for a memory wall, and instead argue that the real limitation -- other than the cost -- lies in the training duration. In addition to this analysis, we introduce two new methods, \textit{layered gradient accumulation} and \textit{modular pipeline parallelism}, which together cut the shortest training time by half. The methods also reduce data movement, lowering the network requirement to a point where a fast InfiniBand connection is not necessary. This increased network efficiency also improve on the methods introduced with the ZeRO optimizer, reducing the memory usage to a tiny fraction of the available GPU memory.

arXiv:2106.01441

Optimization of Heterogeneous Systems with AI Planning Heuristics and Machine Learning: A Performance and Energy Aware Approach

Heterogeneous computing systems provide high performance and energy efficiency. However, to optimally utilize such systems, solutions that distribute the work across host CPUs and accelerating devices are needed. In this paper, we present a performance and energy aware approach that combines AI planning heuristics for parameter space exploration with a machine learning model for performance and energy evaluation to determine a near-optimal system configuration. For data-parallel applications our approach determines a near-optimal host-device distribution of work, number of processing units required and the corresponding scheduling strategy. We evaluate our approach for various heterogeneous systems accelerated with GPU or the Intel Xeon Phi. The experimental results demonstrate that our approach finds a near-optimal system configuration by evaluating only about 7% of reasonable configurations. Furthermore, the performance per Joule estimation of system configurations using our machine learning model is more than 1000x faster compared to the system evaluation by program execution.

arXiv:2106.00992

NVC-Net: End-to-End Adversarial Voice Conversion

Voice conversion has gained increasing popularity in many applications of speech synthesis. The idea is to change the voice identity from one speaker into another while keeping the linguistic content unchanged. Many voice conversion approaches rely on the use of a vocoder to reconstruct the speech from acoustic features, and as a consequence, the speech quality heavily depends on such a vocoder. In this paper, we propose NVC-Net, an end-to-end adversarial network, which performs voice conversion directly on the raw audio waveform of arbitrary length. By disentangling the speaker identity from the speech content, NVC-Net is able to perform non-parallel traditional many-to-many voice conversion as well as zero-shot voice conversion from a short utterance of an unseen target speaker. Importantly, NVC-Net is non-autoregressive and fully convolutional, achieving fast inference. Our model is capable of producing samples at a rate of more than 3600 kHz on an NVIDIA V100 GPU, being orders of magnitude faster than state-of-the-art methods under the same hardware configurations. Objective and subjective evaluations on non-parallel many-to-many voice conversion tasks show that NVC-Net obtains competitive results with significantly fewer parameters.

arXiv:2106.00003

Parallelized Computation and Backpropagation Under Angle-Parametrized Orthogonal Matrices

We present a methodology for parallel acceleration of learning in the presence of matrix orthogonality and unitarity constraints of interest in several branches of machine learning. We show how an apparently sequential elementary rotation parametrization can be restructured into blocks of commutative operations using a well-known tool for coloring the edges of complete graphs, in turn widely applied to schedule round-robin (all-against-all) sports tournaments. The resulting decomposition admits an algorithm to compute a fully-parametrized orthogonal matrix from its rotation parameters in $O(n)$ sequential steps and one to compute the gradient of a training loss with respect to its parameters in $O(n\log n)$ steps. We discuss parametric restrictions of interest to generative modeling and present promising performance results with a prototype GPU implementation.

arXiv:2106.00371

Markov Localisation using Heatmap Regression and Deep Convolutional Odometry

In the context of self-driving vehicles there is strong competition between approaches based on visual localisation and LiDAR. While LiDAR provides important depth information, it is sparse in resolution and expensive. On the other hand, cameras are low-cost and recent developments in deep learning mean they can provide high localisation performance. However, several fundamental problems remain, particularly in the domain of uncertainty, where learning based approaches can be notoriously over-confident. Markov, or grid-based, localisation was an early solution to the localisation problem but fell out of favour due to its computational complexity. Representing the likelihood field as a grid (or volume) means there is a trade off between accuracy and memory size. Furthermore, it is necessary to perform expensive convolutions across the entire likelihood volume. Despite the benefit of simultaneously maintaining a likelihood for all possible locations, grid based approaches were superseded by more efficient particle filters and Monte Carlo Localisation (MCL). However, MCL introduces its own problems e.g. particle deprivation. Recent advances in deep learning hardware allow large likelihood volumes to be stored directly on the GPU, along with the hardware necessary to efficiently perform GPU-bound 3D convolutions and this obviates many of the disadvantages of grid based methods. In this work, we present a novel CNN-based localisation approach that can leverage modern deep learning hardware. By implementing a grid-based Markov localisation approach directly on the GPU, we create a hybrid CNN that can perform image-based localisation and odometry-based likelihood propagation within a single neural network. The resulting approach is capable of outperforming direct pose regression methods as well as state-of-the-art localisation systems.

arXiv:2106.00221

Concurrent Adversarial Learning for Large-Batch Training

Large-batch training has become a commonly used technique when training neural networks with a large number of GPU/TPU processors. As batch size increases, stochastic optimizers tend to converge to sharp local minima, leading to degraded test performance. Current methods usually use extensive data augmentation to increase the batch size, but we found the performance gain with data augmentation decreases as batch size increases, and data augmentation will become insufficient after certain point. In this paper, we propose to use adversarial learning to increase the batch size in large-batch training. Despite being a natural choice for smoothing the decision surface and biasing towards a flat region, adversarial learning has not been successfully applied in large-batch training since it requires at least two sequential gradient computations at each step, which will at least double the running time compared with vanilla training even with a large number of processors. To overcome this issue, we propose a novel Concurrent Adversarial Learning (ConAdv) method that decouple the sequential gradient computations in adversarial learning by utilizing staled parameters. Experimental results demonstrate that ConAdv can successfully increase the batch size on ResNet-50 training on ImageNet while maintaining high accuracy. In particular, we show ConAdv along can achieve 75.3\% top-1 accuracy on ImageNet ResNet-50 training with 96K batch size, and the accuracy can be further improved to 76.2\% when combining ConAdv with data augmentation. This is the first work successfully scales ResNet-50 training batch size to 96K.

arXiv:2106.00186공식 코드 · Papers with Code

Towards Light-weight and Real-time Line Segment Detection

Previous deep learning-based line segment detection (LSD) suffers from the immense model size and high computational cost for line prediction. This constrains them from real-time inference on computationally restricted environments. In this paper, we propose a real-time and light-weight line segment detector for resource-constrained environments named Mobile LSD (M-LSD). We design an extremely efficient LSD architecture by minimizing the backbone network and removing the typical multi-module process for line prediction found in previous methods. To maintain competitive performance with a light-weight network, we present novel training schemes: Segments of Line segment (SoL) augmentation, matching and geometric loss. SoL augmentation splits a line segment into multiple subparts, which are used to provide auxiliary line data during the training process. Moreover, the matching and geometric loss allow a model to capture additional geometric cues. Compared with TP-LSD-Lite, previously the best real-time LSD method, our model (M-LSD-tiny) achieves competitive performance with 2.5% of model size and an increase of 130.5% in inference speed on GPU. Furthermore, our model runs at 56.8 FPS and 48.6 FPS on the latest Android and iPhone mobile devices, respectively. To the best of our knowledge, this is the first real-time deep LSD available on mobile devices. Our code is available.

arXiv:2105.14822공식 코드 · Papers with Code

Effective Batching for Recurrent Neural Network Grammars

As a language model that integrates traditional symbolic operations and flexible neural representations, recurrent neural network grammars (RNNGs) have attracted great attention from both scientific and engineering perspectives. However, RNNGs are known to be harder to scale due to the difficulty of batched training. In this paper, we propose effective batching for RNNGs, where every operation is computed in parallel with tensors across multiple sentences. Our PyTorch implementation effectively employs a GPU and achieves x6 speedup compared to the existing C++ DyNet implementation with model-independent auto-batching. Moreover, our batched RNNG also accelerates inference and achieves x20-150 speedup for beam search depending on beam sizes. Finally, we evaluate syntactic generalization performance of the scaled RNNG against the LSTM baseline, based on the large training data of 100M tokens from English Wikipedia and the broad-coverage targeted syntactic evaluation benchmark. Our RNNG implementation is available at https://github.com/aistairc/rnng-pytorch/.

arXiv:2105.14117공식 코드 · Papers with Code

About Explicit Variance Minimization: Training Neural Networks for Medical Imaging With Limited Data Annotations

Self-supervised learning methods for computer vision have demonstrated the effectiveness of pre-training feature representations, resulting in well-generalizing Deep Neural Networks, even if the annotated data are limited. However, representation learning techniques require a significant amount of time for model training, with most of the time spent on precise hyper-parameter optimization and selection of augmentation techniques. We hypothesized that if the annotated dataset has enough morphological diversity to capture the diversity of the general population, as is common in medical imaging due to conserved similarities of tissue morphology, the variance error of the trained model is the dominant component of the Bias-Variance Trade-off. Therefore, we proposed the Variance Aware Training (VAT) method that exploits this data property by introducing the variance error into the model loss function, thereby, explicitly regularizing the model. Additionally, we provided a theoretical formulation and proof of the proposed method to aid interpreting the approach. Our method requires selecting only one hyper-parameter and matching or improving the performance of state-of-the-art self-supervised methods while achieving an order of magnitude reduction in the GPU training time. We validated VAT on three medical imaging datasets from diverse domains and for various learning objectives. These included a Magnetic Resonance Imaging (MRI) dataset for the heart semantic segmentation (MICCAI 2017 ACDC challenge), fundus photography dataset for ordinary regression of diabetic retinopathy progression (Kaggle 2019 APTOS Blindness Detection challenge), and classification of histopathologic scans of lymph node sections (PatchCamelyon dataset). Our code is available at https://github.com/DmitriiShubin/Variance-Aware-Training.