PyTorchKR Papers

최신 논문

총 5,575편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2012.01250공식 코드 · Papers with Code

Chair Segments: A Compact Benchmark for the Study of Object Segmentation

Over the years, datasets and benchmarks have had an outsized influence on the design of novel algorithms. In this paper, we introduce ChairSegments, a novel and compact semi-synthetic dataset for object segmentation. We also show empirical findings in transfer learning that mirror recent findings for image classification. We particularly show that models that are fine-tuned from a pretrained set of weights lie in the same basin of the optimization landscape. ChairSegments consists of a diverse set of prototypical images of chairs with transparent backgrounds composited into a diverse array of backgrounds. We aim for ChairSegments to be the equivalent of the CIFAR-10 dataset but for quickly designing and iterating over novel model architectures for segmentation. On Chair Segments, a U-Net model can be trained to full convergence in only thirty minutes using a single GPU. Finally, while this dataset is semi-synthetic, it can be a useful proxy for real data, leading to state-of-the-art accuracy on the Object Discovery dataset when used as a source of pretraining.

arXiv:2012.01050

Learning Universal Shape Dictionary for Realtime Instance Segmentation

We present a novel explicit shape representation for instance segmentation. Based on how to model the object shape, current instance segmentation systems can be divided into two categories, implicit and explicit models. The implicit methods, which represent the object mask/contour by intractable network parameters, and produce it through pixel-wise classification, are predominant. However, the explicit methods, which parameterize the shape with simple and explainable models, are less explored. Since the operations to generate the final shape are light-weighted, the explicit methods have a clear speed advantage over implicit methods, which is crucial for real-world applications. The proposed USD-Seg adopts a linear model, sparse coding with dictionary, for object shapes. First, it learns a dictionary from a large collection of shape datasets, making any shape being able to be decomposed into a linear combination through the dictionary. Hence the name "Universal Shape Dictionary". Then it adds a simple shape vector regression head to ordinary object detector, giving the detector segmentation ability with minimal overhead. For quantitative evaluation, we use both average precision (AP) and the proposed Efficiency of AP (AP$_E$) metric, which intends to also measure the computational consumption of the framework to cater to the requirements of real-world applications. We report experimental results on the challenging COCO dataset, in which our single model on a single Titan Xp GPU achieves 35.8 AP and 27.8 AP$_E$ at 65 fps with YOLOv4 as base detector, 34.1 AP and 28.6 AP$_E$ at 12 fps with FCOS as base detector.

arXiv:2012.00899

Displacement-Invariant Cost Computation for Efficient Stereo Matching

Although deep learning-based methods have dominated stereo matching leaderboards by yielding unprecedented disparity accuracy, their inference time is typically slow, on the order of seconds for a pair of 540p images. The main reason is that the leading methods employ time-consuming 3D convolutions applied to a 4D feature volume. A common way to speed up the computation is to downsample the feature volume, but this loses high-frequency details. To overcome these challenges, we propose a \emph{displacement-invariant cost computation module} to compute the matching costs without needing a 4D feature volume. Rather, costs are computed by applying the same 2D convolution network on each disparity-shifted feature map pair independently. Unlike previous 2D convolution-based methods that simply perform context mapping between inputs and disparity maps, our proposed approach learns to match features between the two images. We also propose an entropy-based refinement strategy to refine the computed disparity map, which further improves speed by avoiding the need to compute a second disparity map on the right image. Extensive experiments on standard datasets (SceneFlow, KITTI, ETH3D, and Middlebury) demonstrate that our method achieves competitive accuracy with much less inference time. On typical image sizes, our method processes over 100 FPS on a desktop GPU, making our method suitable for time-critical applications such as autonomous driving. We also show that our approach generalizes well to unseen datasets, outperforming 4D-volumetric methods.

arXiv:2012.00365

Python Workflows on HPC Systems

The recent successes and wide spread application of compute intensive machine learning and data analytics methods have been boosting the usage of the Python programming language on HPC systems. While Python provides many advantages for the users, it has not been designed with a focus on multi-user environments or parallel programming - making it quite challenging to maintain stable and secure Python workflows on a HPC system. In this paper, we analyze the key problems induced by the usage of Python on HPC clusters and sketch appropriate workarounds for efficiently maintaining multi-user Python software environments, securing and restricting resources of Python jobs and containing Python processes, while focusing on Deep Learning applications running on GPU clusters.

arXiv:2012.00211

Toward Accurate Platform-Aware Performance Modeling for Deep Neural Networks

In this paper, we provide a fine-grain machine learning-based method, PerfNetV2, which improves the accuracy of our previous work for modeling the neural network performance on a variety of GPU accelerators. Given an application, the proposed method can be used to predict the inference time and training time of the convolutional neural networks used in the application, which enables the system developer to optimize the performance by choosing the neural networks and/or incorporating the hardware accelerators to deliver satisfactory results in time. Furthermore, the proposed method is capable of predicting the performance of an unseen or non-existing device, e.g. a new GPU which has a higher operating frequency with less processor cores, but more memory capacity. This allows a system developer to quickly search the hardware design space and/or fine-tune the system configuration. Compared to the previous works, PerfNetV2 delivers more accurate results by modeling detailed host-accelerator interactions in executing the full neural networks and improving the architecture of the machine learning model used in the predictor. Our case studies show that PerfNetV2 yields a mean absolute percentage error within 13.1% on LeNet, AlexNet, and VGG16 on NVIDIA GTX-1080Ti, while the error rate on a previous work published in ICBD 2018 could be as large as 200%.

arXiv:2011.14203

EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP Inference

Transformer-based language models such as BERT provide significant accuracy improvement for a multitude of natural language processing (NLP) tasks. However, their hefty computational and memory demands make them challenging to deploy to resource-constrained edge platforms with strict latency requirements. We present EdgeBERT, an in-depth algorithm-hardware co-design for latency-aware energy optimization for multi-task NLP. EdgeBERT employs entropy-based early exit predication in order to perform dynamic voltage-frequency scaling (DVFS), at a sentence granularity, for minimal energy consumption while adhering to a prescribed target latency. Computation and memory footprint overheads are further alleviated by employing a calibrated combination of adaptive attention span, selective network pruning, and floating-point quantization. Furthermore, in order to maximize the synergistic benefits of these algorithms in always-on and intermediate edge computing settings, we specialize a 12nm scalable hardware accelerator system, integrating a fast-switching low-dropout voltage regulator (LDO), an all-digital phase-locked loop (ADPLL), as well as, high-density embedded non-volatile memories (eNVMs) wherein the sparse floating-point bit encodings of the shared multi-task parameters are carefully stored. Altogether, latency-aware multi-task NLP inference acceleration on the EdgeBERT hardware system generates up to 7x, 2.5x, and 53x lower energy compared to the conventional inference without early stopping, the latency-unbounded early exit approach, and CUDA adaptations on an Nvidia Jetson Tegra X2 mobile GPU, respectively.

arXiv:2011.14133

Towards Fast and Light-Weight Restoration of Dark Images

The ability to capture good quality images in the dark and near-zero lux conditions has been a long-standing pursuit of the computer vision community. The seminal work by Chen et al. [5] has especially caused renewed interest in this area, resulting in methods that build on top of their work in a bid to improve the reconstruction. However, for practical utility and deployment of low-light enhancement algorithms on edge devices such as embedded systems, surveillance cameras, autonomous robots and smartphones, the solution must respect additional constraints such as limited GPU memory and processing power. With this in mind, we propose a deep neural network architecture that aims to strike a balance between the network latency, memory utilization, model parameters, and reconstruction quality. The key idea is to forbid computations in the High-Resolution (HR) space and limit them to a Low-Resolution (LR) space. However, doing the bulk of computations in the LR space causes artifacts in the restored image. We thus propose Pack and UnPack operations, which allow us to effectively transit between the HR and LR spaces without incurring much artifacts in the restored image. We show that we can enhance a full resolution, 2848 x 4256, extremely dark single-image in the ballpark of 3 seconds even on a CPU. We achieve this with 2 - 7x fewer model parameters, 2 - 3x lower memory utilization, 5 - 20x speed up and yet maintain a competitive image reconstruction quality compared to the state-of-the-art algorithms.

arXiv:2011.14773

Deep learning approach to left ventricular non-compaction measurement

Left ventricular non-compaction (LVNC) is a rare cardiomyopathy characterized by abnormal trabeculations in the left ventricle cavity. Although traditional computer vision approaches exist for LVNC diagnosis, deep learning-based tools could not be found in the literature. In this paper, a first approach using convolutional neural networks (CNNs) is presented. Four CNNs are trained to automatically segment the compacted and trabecular areas of the left ventricle for a population of patients diagnosed with Hypertrophic cardiomyopathy. Inference results confirm that deep learning-based approaches can achieve excellent results in the diagnosis and measurement of LVNC. The two best CNNs (U-Net and Efficient U-Net B1) perform image segmentation in less than 0.2 s on a CPU and in less than 0.01 s on a GPU. Additionally, a subjective evaluation of the output images with the identified zones is performed by expert cardiologists, with a perfect visual agreement for all the slices, outperforming already existing automatic tools.

arXiv:2011.14959

Deep Dose Plugin Towards Real-time Monte Carlo Dose Calculation Through a Deep Learning based Denoising Algorithm

Monte Carlo (MC) simulation is considered the gold standard method for radiotherapy dose calculation. However, achieving high precision requires a large number of simulation histories, which is time consuming. The use of computer graphics processing units (GPUs) has greatly accelerated MC simulation and allows dose calculation within a few minutes for a typical radiotherapy treatment plan. However, some clinical applications demand real time efficiency for MC dose calculation. To tackle this problem, we have developed a real time, deep learning based dose denoiser that can be plugged into a current GPU based MC dose engine to enable real time MC dose calculation. We used two different acceleration strategies to achieve this goal: 1) we applied voxel unshuffle and voxel shuffle operators to decrease the input and output sizes without any information loss, and 2) we decoupled the 3D volumetric convolution into a 2D axial convolution and a 1D slice convolution. In addition, we used a weakly supervised learning framework to train the network, which greatly reduces the size of the required training dataset and thus enables fast fine tuning based adaptation of the trained model to different radiation beams. Experimental results show that the proposed denoiser can run in as little as 39 ms, which is around 11.6 times faster than the baseline model. As a result, the whole MC dose calculation pipeline can be finished within 0.15 seconds, including both GPU MC dose calculation and deep learning based denoising, achieving the real time efficiency needed for some radiotherapy applications, such as online adaptive radiotherapy.

arXiv:2011.14439공식 코드 · Papers with Code

Scaling Down Deep Learning with MNIST-1D

Although deep learning models have taken on commercial and political relevance, key aspects of their training and operation remain poorly understood. This has sparked interest in science of deep learning projects, many of which require large amounts of time, money, and electricity. But how much of this research really needs to occur at scale? In this paper, we introduce MNIST-1D: a minimalist, procedurally generated, low-memory, and low-compute alternative to classic deep learning benchmarks. Although the dimensionality of MNIST-1D is only 40 and its default training set size only 4000, MNIST-1D can be used to study inductive biases of different deep architectures, find lottery tickets, observe deep double descent, metalearn an activation function, and demonstrate guillotine regularization in self-supervised learning. All these experiments can be conducted on a GPU or often even on a CPU within minutes, allowing for fast prototyping, educational use cases, and cutting-edge research on a low budget.

arXiv:2011.13824공식 코드 · Papers with Code

Fast and Complete: Enabling Complete Neural Network Verification with Rapid and Massively Parallel Incomplete Verifiers

Formal verification of neural networks (NNs) is a challenging and important problem. Existing efficient complete solvers typically require the branch-and-bound (BaB) process, which splits the problem domain into sub-domains and solves each sub-domain using faster but weaker incomplete verifiers, such as Linear Programming (LP) on linearly relaxed sub-domains. In this paper, we propose to use the backward mode linear relaxation based perturbation analysis (LiRPA) to replace LP during the BaB process, which can be efficiently implemented on the typical machine learning accelerators such as GPUs and TPUs. However, unlike LP, LiRPA when applied naively can produce much weaker bounds and even cannot check certain conflicts of sub-domains during splitting, making the entire procedure incomplete after BaB. To address these challenges, we apply a fast gradient based bound tightening procedure combined with batch splits and the design of minimal usage of LP bound procedure, enabling us to effectively use LiRPA on the accelerator hardware for the challenging complete NN verification problem and significantly outperform LP-based approaches. On a single GPU, we demonstrate an order of magnitude speedup compared to existing LP-based approaches.

arXiv:2011.13591

Multi-objective Neural Architecture Search with Almost No Training

In the recent past, neural architecture search (NAS) has attracted increasing attention from both academia and industries. Despite the steady stream of impressive empirical results, most existing NAS algorithms are computationally prohibitive to execute due to the costly iterations of stochastic gradient descent (SGD) training. In this work, we propose an effective alternative, dubbed Random-Weight Evaluation (RWE), to rapidly estimate the performance of network architectures. By just training the last linear classification layer, RWE reduces the computational cost of evaluating an architecture from hours to seconds. When integrated within an evolutionary multi-objective algorithm, RWE obtains a set of efficient architectures with state-of-the-art performance on CIFAR-10 with less than two hours' searching on a single GPU card. Ablation studies on rank-order correlations and transfer learning experiments to ImageNet have further validated the effectiveness of RWE.

arXiv:2011.12490

DeRF: Decomposed Radiance Fields

With the advent of Neural Radiance Fields (NeRF), neural networks can now render novel views of a 3D scene with quality that fools the human eye. Yet, generating these images is very computationally intensive, limiting their applicability in practical scenarios. In this paper, we propose a technique based on spatial decomposition capable of mitigating this issue. Our key observation is that there are diminishing returns in employing larger (deeper and/or wider) networks. Hence, we propose to spatially decompose a scene and dedicate smaller networks for each decomposed part. When working together, these networks can render the whole scene. This allows us near-constant inference time regardless of the number of decomposed parts. Moreover, we show that a Voronoi spatial decomposition is preferable for this purpose, as it is provably compatible with the Painter's Algorithm for efficient and GPU-friendly rendering. Our experiments show that for real-world scenes, our method provides up to 3x more efficient inference than NeRF (with the same rendering quality), or an improvement of up to 1.0~dB in PSNR (for the same inference cost).

arXiv:2011.12790공식 코드 · Papers with Code

Fast Region Proposal Learning for Object Detection for Robotics

Object detection is a fundamental task for robots to operate in unstructured environments. Today, there are several deep learning algorithms that solve this task with remarkable performance. Unfortunately, training such systems requires several hours of GPU time. For robots, to successfully adapt to changes in the environment or learning new objects, it is also important that object detectors can be re-trained in a short amount of time. A recent method [1] proposes an architecture that leverages on the powerful representation of deep learning descriptors, while permitting fast adaptation time. Leveraging on the natural decomposition of the task in (i) regions candidate generation, (ii) feature extraction and (iii) regions classification, this method performs fast adaptation of the detector, by only re-training the classification layer. This shortens training time while maintaining state-of-the-art performance. In this paper, we firstly demonstrate that a further boost in accuracy can be obtained by adapting, in addition, the regions candidate generation on the task at hand. Secondly, we extend the object detection system presented in [1] with the proposed fast learning approach, showing experimental evidence on the improvement provided in terms of speed and accuracy on two different robotics datasets. The code to reproduce the experiments is publicly available on GitHub.

arXiv:2011.11890공식 코드 · Papers with Code

Cross-Camera Convolutional Color Constancy

We present "Cross-Camera Convolutional Color Constancy" (C5), a learning-based method, trained on images from multiple cameras, that accurately estimates a scene's illuminant color from raw images captured by a new camera previously unseen during training. C5 is a hypernetwork-like extension of the convolutional color constancy (CCC) approach: C5 learns to generate the weights of a CCC model that is then evaluated on the input image, with the CCC weights dynamically adapted to different input content. Unlike prior cross-camera color constancy models, which are usually designed to be agnostic to the spectral properties of test-set images from unobserved cameras, C5 approaches this problem through the lens of transductive inference: additional unlabeled images are provided as input to the model at test time, which allows the model to calibrate itself to the spectral properties of the test-set camera during inference. C5 achieves state-of-the-art accuracy for cross-camera color constancy on several datasets, is fast to evaluate (~7 and ~90 ms per image on a GPU or CPU, respectively), and requires little memory (~2 MB), and thus is a practical solution to the problem of calibration-free automatic white balance for mobile photography.

arXiv:2011.11961공식 코드 · Papers with Code

MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition

Existing portrait matting methods either require auxiliary inputs that are costly to obtain or involve multiple stages that are computationally expensive, making them less suitable for real-time applications. In this work, we present a light-weight matting objective decomposition network (MODNet) for portrait matting in real-time with a single input image. The key idea behind our efficient design is by optimizing a series of sub-objectives simultaneously via explicit constraints. In addition, MODNet includes two novel techniques for improving model efficiency and robustness. First, an Efficient Atrous Spatial Pyramid Pooling (e-ASPP) module is introduced to fuse multi-scale features for semantic estimation. Second, a self-supervised sub-objectives consistency (SOC) strategy is proposed to adapt MODNet to real-world data to address the domain shift problem common to trimap-free methods. MODNet is easy to be trained in an end-to-end manner. It is much faster than contemporaneous methods and runs at 67 frames per second on a 1080Ti GPU. Experiments show that MODNet outperforms prior trimap-free methods by a large margin on both Adobe Matting Dataset and a carefully designed photographic portrait matting (PPM-100) benchmark proposed by us. Further, MODNet achieves remarkable results on daily photos and videos. Our code and models are available at https://github.com/ZHKKKe/MODNet, and the PPM-100 benchmark is released at https://github.com/ZHKKKe/PPM.

arXiv:2011.11134공식 코드 · Papers with Code

Differentiable Computational Geometry for 2D and 3D machine learning

With the growth of machine learning algorithms with geometry primitives, a high-efficiency library with differentiable geometric operators are desired. We present an optimized Differentiable Geometry Algorithm Library (DGAL) loaded with implementations of differentiable operators for geometric primitives like lines and polygons. The library is a header-only templated C++ library with GPU support. We discuss the internal design of the library and benchmark its performance on some tasks with other implementations.

arXiv:2011.10912

Third ArchEdge Workshop: Exploring the Design Space of Efficient Deep Neural Networks

This paper gives an overview of our ongoing work on the design space exploration of efficient deep neural networks (DNNs). Specifically, we cover two aspects: (1) static architecture design efficiency and (2) dynamic model execution efficiency. For static architecture design, different from existing end-to-end hardware modeling assumptions, we conduct full-stack profiling at the GPU core level to identify better accuracy-latency trade-offs for DNN designs. For dynamic model execution, different from prior work that tackles model redundancy at the DNN-channels level, we explore a new dimension of DNN feature map redundancy to be dynamically traversed at runtime. Last, we highlight several open questions that are poised to draw research attention in the next few years.

arXiv:2011.10251

On-Device Text Image Super Resolution

Recent research on super-resolution (SR) has witnessed major developments with the advancements of deep convolutional neural networks. There is a need for information extraction from scenic text images or even document images on device, most of which are low-resolution (LR) images. Therefore, SR becomes an essential pre-processing step as Bicubic Upsampling, which is conventionally present in smartphones, performs poorly on LR images. To give the user more control over his privacy, and to reduce the carbon footprint by reducing the overhead of cloud computing and hours of GPU usage, executing SR models on the edge is a necessity in the recent times. There are various challenges in running and optimizing a model on resource-constrained platforms like smartphones. In this paper, we present a novel deep neural network that reconstructs sharper character edges and thus boosts OCR confidence. The proposed architecture not only achieves significant improvement in PSNR over bicubic upsampling on various benchmark datasets but also runs with an average inference time of 11.7 ms per image. We have outperformed state-of-the-art on the Text330 dataset. We also achieve an OCR accuracy of 75.89% on the ICDAR 2015 TextSR dataset, where ground truth has an accuracy of 78.10%.

arXiv:2011.09214

Res-GCNN: A Lightweight Residual Graph Convolutional Neural Networks for Human Trajectory Forecasting

Autonomous driving vehicles (ADVs) hold great hopes to solve traffic congestion problems and reduce the number of traffic accidents. Accurate trajectories prediction of other traffic agents around ADVs is of key importance to achieve safe and efficient driving. Pedestrians, particularly, are more challenging to forecast due to their complex social in-teractions and randomly moving patterns. We propose a Residual Graph Convolutional Neural Network (Res-GCNN), which models the interactive behaviors of pedes-trians by using the adjacent matrix of the constructed graph for the current scene. Though the proposed Res-GCNN is quite lightweight with only about 6.4 kilo parameters which outperforms all other methods in terms of parameters size, our experimental results show an improvement over the state of art by 13.3% on the Final Displacement Error (FDE) which reaches 0.65 meter. As for the Average Dis-placement Error (ADE), we achieve a suboptimal result (the value is 0.37 meter), which is also very competitive. The Res-GCNN is evaluated in the platform with an NVIDIA GeForce RTX1080Ti GPU, and its mean inference time of the whole dataset is only about 2.2 microseconds. Compared with other methods, the proposed method shows strong potential for onboard application accounting for forecasting accuracy and time efficiency. The code will be made publicly available on GitHub.