PyTorchKR Papers

최신 논문

총 5,575편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2006.06762

Ansor: Generating High-Performance Tensor Programs for Deep Learning

High-performance tensor programs are crucial to guarantee efficient execution of deep neural networks. However, obtaining performant tensor programs for different operators on various hardware platforms is notoriously challenging. Currently, deep learning systems rely on vendor-provided kernel libraries or various search strategies to get performant tensor programs. These approaches either require significant engineering effort to develop platform-specific optimization code or fall short of finding high-performance programs due to restricted search space and ineffective exploration strategy. We present Ansor, a tensor program generation framework for deep learning applications. Compared with existing search strategies, Ansor explores many more optimization combinations by sampling programs from a hierarchical representation of the search space. Ansor then fine-tunes the sampled programs with evolutionary search and a learned cost model to identify the best programs. Ansor can find high-performance programs that are outside the search space of existing state-of-the-art approaches. In addition, Ansor utilizes a task scheduler to simultaneously optimize multiple subgraphs in deep neural networks. We show that Ansor improves the execution performance of deep neural networks relative to the state-of-the-art on the Intel CPU, ARM CPU, and NVIDIA GPU by up to $3.8\times$, $2.6\times$, and $1.7\times$, respectively.

arXiv:2006.06185공식 코드 · Papers with Code

JIT-Masker: Efficient Online Distillation for Background Matting

We design a real-time portrait matting pipeline for everyday use, particularly for "virtual backgrounds" in video conferences. Existing segmentation and matting methods prioritize accuracy and quality over throughput and efficiency, and our pipeline enables trading off a controllable amount of accuracy for better throughput by leveraging online distillation on the input video stream. We construct our own dataset of simulated video calls in various scenarios, and show that our approach delivers a 5x speedup over a saliency detection based pipeline in a non-GPU accelerated setting while delivering higher quality results. We demonstrate that an online distillation approach can feasibly work as part of a general, consumer level product as a "virtual background" tool. Our public implementation is at https://github.com/josephch405/jit-masker.

arXiv:2006.05075

Artificial Intelligence (AI)-Centric Management of Resources in Modern Distributed Computing Systems

Contemporary Distributed Computing Systems (DCS) such as Cloud Data Centres are large scale, complex, heterogeneous, and distributed across multiple networks and geographical boundaries. On the other hand, the Internet of Things (IoT)-driven applications are producing a huge amount of data that requires real-time processing and fast response. Managing these resources efficiently to provide reliable services to end-users or applications is a challenging task. The existing Resource Management Systems (RMS) rely on either static or heuristic solutions inadequate for such composite and dynamic systems. The advent of Artificial Intelligence (AI) due to data availability and processing capabilities manifested into possibilities of exploring data-driven solutions in RMS tasks that are adaptive, accurate, and efficient. In this regard, this paper aims to draw the motivations and necessities for data-driven solutions in resource management. It identifies the challenges associated with it and outlines the potential future research directions detailing where and how to apply the data-driven techniques in the different RMS tasks. Finally, it provides a conceptual data-driven RMS model for DCS and presents the two real-time use cases (GPU frequency scaling and data centre resource management from Google Cloud and Microsoft Azure) demonstrating AI-centric approaches' feasibility.

arXiv:2006.05325

ComboNet: Combined 2D & 3D Architecture for Aorta Segmentation

3D segmentation with deep learning if trained with full resolution is the ideal way of achieving the best accuracy. Unlike in 2D, 3D segmentation generally does not have sparse outliers, prevents leakage to surrounding soft tissues, at the very least it is generally more consistent than 2D segmentation. However, GPU memory is generally the bottleneck for such an application. Thus, most of the 3D segmentation applications handle sub-sampled input instead of full resolution, which comes with the cost of losing precision at the boundary. In order to maintain precision at the boundary and prevent sparse outliers and leakage, we designed ComboNet. ComboNet is designed in an end to end fashion with three sub-network structures. The first two are parallel: 2D UNet with full resolution and 3D UNet with four times sub-sampled input. The last stage is the concatenation of 2D and 3D outputs along with a full-resolution input image which is followed by two convolution layers either with 2D or 3D convolutions. With ComboNet we have achieved $92.1\%$ dice accuracy for aorta segmentation. With Combonet, we have observed up to $2.3\%$ improvement of dice accuracy as opposed to 2D UNet with the full-resolution input image.

arXiv:2006.05183

A Note on Deepfake Detection with Low-Resources

Deepfakes are videos that include changes, quite often substituting face of a portrayed individual with a different face using neural networks. Even though the technology gained its popularity as a carrier of jokes and parodies it raises a serious threat to ones security - via biometric impersonation or besmearing. In this paper we present two methods that allow detecting Deepfakes for a user without significant computational power. In particular, we enhance MesoNet by replacing the original activation functions allowing a nearly 1% improvement as well as increasing the consistency of the results. Moreover, we introduced and verified a new activation function - Pish that at the cost of slight time overhead allows even higher consistency. Additionally, we present a preliminary results of Deepfake detection method based on Local Feature Descriptors (LFD), that allows setting up the system even faster and without resorting to GPU computation. Our method achieved Equal Error Rate of 0.28, with both accuracy and recall exceeding 0.7.

arXiv:2006.05117공식 코드 · Papers with Code

Hysia: Serving DNN-Based Video-to-Retail Applications in Cloud

Combining \underline{v}ideo streaming and online \underline{r}etailing (V2R) has been a growing trend recently. In this paper, we provide practitioners and researchers in multimedia with a cloud-based platform named Hysia for easy development and deployment of V2R applications. The system consists of: 1) a back-end infrastructure providing optimized V2R related services including data engine, model repository, model serving and content matching; and 2) an application layer which enables rapid V2R application prototyping. Hysia addresses industry and academic needs in large-scale multimedia by: 1) seamlessly integrating state-of-the-art libraries including NVIDIA video SDK, Facebook faiss, and gRPC; 2) efficiently utilizing GPU computation; and 3) allowing developers to bind new models easily to meet the rapidly changing deep learning (DL) techniques. On top of that, we implement an orchestrator for further optimizing DL model serving performance. Hysia has been released as an open source project on GitHub, and attracted considerable attention. We have published Hysia to DockerHub as an official image for seamless integration and deployment in current cloud environments.

arXiv:2006.05518

MVLidarNet: Real-Time Multi-Class Scene Understanding for Autonomous Driving Using Multiple Views

Autonomous driving requires the inference of actionable information such as detecting and classifying objects, and determining the drivable space. To this end, we present Multi-View LidarNet (MVLidarNet), a two-stage deep neural network for multi-class object detection and drivable space segmentation using multiple views of a single LiDAR point cloud. The first stage processes the point cloud projected onto a perspective view in order to semantically segment the scene. The second stage then processes the point cloud (along with semantic labels from the first stage) projected onto a bird's eye view, to detect and classify objects. Both stages use an encoder-decoder architecture. We show that our multi-view, multi-stage, multi-class approach is able to detect and classify objects while simultaneously determining the drivable space using a single LiDAR scan as input, in challenging scenes with more than one hundred vehicles and pedestrians at a time. The system operates efficiently at 150 fps on an embedded GPU designed for a self-driving car, including a postprocessing step to maintain identities over time. We show results on both KITTI and a much larger internal dataset, thus demonstrating the method's ability to scale by an order of magnitude.

arXiv:2006.04388공식 코드 · Papers with Code

Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

One-stage detector basically formulates object detection as dense classification and localization. The classification is usually optimized by Focal Loss and the box location is commonly learned under Dirac delta distribution. A recent trend for one-stage detectors is to introduce an individual prediction branch to estimate the quality of localization, where the predicted quality facilitates the classification to improve detection performance. This paper delves into the representations of the above three fundamental elements: quality estimation, classification and localization. Two problems are discovered in existing practices, including (1) the inconsistent usage of the quality estimation and classification between training and inference and (2) the inflexible Dirac delta distribution for localization when there is ambiguity and uncertainty in complex scenes. To address the problems, we design new representations for these elements. Specifically, we merge the quality estimation into the class prediction vector to form a joint representation of localization quality and classification, and use a vector to represent arbitrary distribution of box locations. The improved representations eliminate the inconsistency risk and accurately depict the flexible distribution in real data, but contain continuous labels, which is beyond the scope of Focal Loss. We then propose Generalized Focal Loss (GFL) that generalizes Focal Loss from its discrete form to the continuous version for successful optimization. On COCO test-dev, GFL achieves 45.0\% AP using ResNet-101 backbone, surpassing state-of-the-art SAPD (43.5\%) and ATSS (43.6\%) with higher or comparable inference speed, under the same backbone and training settings. Notably, our best model can achieve a single-model single-scale AP of 48.2\%, at 10 FPS on a single 2080Ti GPU. Code and models are available at https://github.com/implus/GFocal.

arXiv:2006.04647공식 코드 · Papers with Code

Neural Architecture Search without Training

The time and effort involved in hand-designing deep neural networks is immense. This has prompted the development of Neural Architecture Search (NAS) techniques to automate this design. However, NAS algorithms tend to be slow and expensive; they need to train vast numbers of candidate networks to inform the search process. This could be alleviated if we could partially predict a network's trained accuracy from its initial state. In this work, we examine the overlap of activations between datapoints in untrained networks and motivate how this can give a measure which is usefully indicative of a network's trained performance. We incorporate this measure into a simple algorithm that allows us to search for powerful networks without any training in a matter of seconds on a single GPU, and verify its effectiveness on NAS-Bench-101, NAS-Bench-201, NATS-Bench, and Network Design Spaces. Our approach can be readily combined with more expensive search methods; we examine a simple adaptation of regularised evolutionary search. Code for reproducing our experiments is available at https://github.com/BayesWatch/nas-without-training.

arXiv:2006.04593공식 코드 · Papers with Code

ARIANN: Low-Interaction Privacy-Preserving Deep Learning via Function Secret Sharing

We propose AriaNN, a low-interaction privacy-preserving framework for private neural network training and inference on sensitive data. Our semi-honest 2-party computation protocol (with a trusted dealer) leverages function secret sharing, a recent lightweight cryptographic protocol that allows us to achieve an efficient online phase. We design optimized primitives for the building blocks of neural networks such as ReLU, MaxPool and BatchNorm. For instance, we perform private comparison for ReLU operations with a single message of the size of the input during the online phase, and with preprocessing keys close to 4X smaller than previous work. Last, we propose an extension to support n-party private federated learning. We implement our framework as an extensible system on top of PyTorch that leverages CPU and GPU hardware acceleration for cryptographic and machine learning operations. We evaluate our end-to-end system for private inference between distant servers on standard neural networks such as AlexNet, VGG16 or ResNet18, and for private training on smaller networks like LeNet. We show that computation rather than communication is the main bottleneck and that using GPUs together with reduced key size is a promising solution to overcome this barrier.

arXiv:2006.03394공식 코드 · Papers with Code

Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels

Prostate cancer is the most prevalent cancer among men in Western countries, with 1.1 million new diagnoses every year. The gold standard for the diagnosis of prostate cancer is a pathologists' evaluation of prostate tissue. To potentially assist pathologists deep-learning-based cancer detection systems have been developed. Many of the state-of-the-art models are patch-based convolutional neural networks, as the use of entire scanned slides is hampered by memory limitations on accelerator cards. Patch-based systems typically require detailed, pixel-level annotations for effective training. However, such annotations are seldom readily available, in contrast to the clinical reports of pathologists, which contain slide-level labels. As such, developing algorithms which do not require manual pixel-wise annotations, but can learn using only the clinical report would be a significant advancement for the field. In this paper, we propose to use a streaming implementation of convolutional layers, to train a modern CNN (ResNet-34) with 21 million parameters end-to-end on 4712 prostate biopsies. The method enables the use of entire biopsy images at high-resolution directly by reducing the GPU memory requirements by 2.4 TB. We show that modern CNNs, trained using our streaming approach, can extract meaningful features from high-resolution images without additional heuristics, reaching similar performance as state-of-the-art patch-based and multiple-instance learning methods. By circumventing the need for manual annotations, this approach can function as a blueprint for other tasks in histopathological diagnosis. The source code to reproduce the streaming models is available at https://github.com/DIAGNijmegen/pathology-streaming-pipeline .

arXiv:2006.03250

FP-Stereo: Hardware-Efficient Stereo Vision for Embedded Applications

Fast and accurate depth estimation, or stereo matching, is essential in embedded stereo vision systems, requiring substantial design effort to achieve an appropriate balance among accuracy, speed and hardware cost. To reduce the design effort and achieve the right balance, we propose FP-Stereo for building high-performance stereo matching pipelines on FPGAs automatically. FP-Stereo consists of an open-source hardware-efficient library, allowing designers to obtain the desired implementation instantly. Diverse methods are supported in our library for each stage of the stereo matching pipeline and a series of techniques are developed to exploit the parallelism and reduce the resource overhead. To improve the usability, FP-Stereo can generate synthesizable C code of the FPGA accelerator with our optimized HLS templates automatically. To guide users for the right design choice meeting specific application requirements, detailed comparisons are performed on various configurations of our library to investigate the accuracy/speed/cost trade-off. Experimental results also show that FP-Stereo outperforms the state-of-the-art FPGA design from all aspects, including 6.08% lower error, 2x faster speed, 30% less resource usage and 40% less energy consumption. Compared to GPU designs, FP-Stereo achieves the same accuracy at a competitive speed while consuming much less energy.

arXiv:2006.02631공식 코드 · Papers with Code

FastReID: A Pytorch Toolbox for General Instance Re-identification

General Instance Re-identification is a very important task in the computer vision, which can be widely used in many practical applications, such as person/vehicle re-identification, face recognition, wildlife protection, commodity tracing, and snapshop, etc.. To meet the increasing application demand for general instance re-identification, we present FastReID as a widely used software system in JD AI Research. In FastReID, highly modular and extensible design makes it easy for the researcher to achieve new research ideas. Friendly manageable system configuration and engineering deployment functions allow practitioners to quickly deploy models into productions. We have implemented some state-of-the-art projects, including person re-id, partial re-id, cross-domain re-id and vehicle re-id, and plan to release these pre-trained models on multiple benchmark datasets. FastReID is by far the most general and high-performance toolbox that supports single and multiple GPU servers, you can reproduce our project results very easily and are very welcome to use it, the code and models are available at https://github.com/JDAI-CV/fast-reid.

arXiv:2006.02834

Look Locally Infer Globally: A Generalizable Face Anti-Spoofing Approach

State-of-the-art spoof detection methods tend to overfit to the spoof types seen during training and fail to generalize to unknown spoof types. Given that face anti-spoofing is inherently a local task, we propose a face anti-spoofing framework, namely Self-Supervised Regional Fully Convolutional Network (SSR-FCN), that is trained to learn local discriminative cues from a face image in a self-supervised manner. The proposed framework improves generalizability while maintaining the computational efficiency of holistic face anti-spoofing approaches (< 4 ms on a Nvidia GTX 1080Ti GPU). The proposed method is interpretable since it localizes which parts of the face are labeled as spoofs. Experimental results show that SSR-FCN can achieve TDR = 65% @ 2.0% FDR when evaluated on a dataset comprising of 13 different spoof types under unknown attacks while achieving competitive performances under standard benchmark datasets (Oulu-NPU, CASIA-MFSD, and Replay-Attack).

arXiv:2006.02267공식 코드 · Papers with Code

FastONN -- Python based open-source GPU implementation for Operational Neural Networks

Operational Neural Networks (ONNs) have recently been proposed as a special class of artificial neural networks for grid structured data. They enable heterogenous non-linear operations to generalize the widely adopted convolution-based neuron model. This work introduces a fast GPU-enabled library for training operational neural networks, FastONN, which is based on a novel vectorized formulation of the operational neurons. Leveraging on automatic reverse-mode differentiation for backpropagation, FastONN enables increased flexibility with the incorporation of new operator sets and customized gradient flows. Additionally, bundled auxiliary modules offer interfaces for performance tracking and checkpointing across different data partitions and customized metrics.

arXiv:2006.01201공식 코드 · Papers with Code

High-quality Panorama Stitching based on Asymmetric Bidirectional Optical Flow

In this paper, we propose a panorama stitching algorithm based on asymmetric bidirectional optical flow. This algorithm expects multiple photos captured by fisheye lens cameras as input, and then, through the proposed algorithm, these photos can be merged into a high-quality 360-degree spherical panoramic image. For photos taken from a distant perspective, the parallax among them is relatively small, and the obtained panoramic image can be nearly seamless and undistorted. For photos taken from a close perspective or with a relatively large parallax, a seamless though partially distorted panoramic image can also be obtained. Besides, with the help of Graphics Processing Unit (GPU), this algorithm can complete the whole stitching process at a very fast speed: typically, it only takes less than 30s to obtain a panoramic image of 9000-by-4000 pixels, which means our panorama stitching algorithm is of high value in many real-time applications. Our code is available at https://github.com/MungoMeng/Panorama-OpticalFlow.

arXiv:2006.01238

SOT-MRAM based Sigmoidal Neuron for Neuromorphic Architectures

In this paper, the intrinsic physical characteristics of spin-orbit torque (SOT) magnetoresistive random-access memory (MRAM) devices are leveraged to realize sigmoidal neurons in neuromorphic architectures. Performance comparisons with the previous power- and area-efficient sigmoidal neuron circuits exhibit 74x and 12x reduction in power-area-product values for the proposed SOT-MRAM based neuron. To verify the functionally of the proposed neuron within larger scale designs, we have implemented a circuit realization of a 784x16x10 SOT-MRAM based multiplayer perceptron (MLP) for MNIST pattern recognition application using SPICE circuit simulation tool. The results obtained exhibit that the proposed SOT-MRAM based MLP can achieve accuracies comparable to an ideal binarized MLP architecture implemented on GPU, while realizing orders of magnitude increase in processing speed.

arXiv:2006.00844

Distilling Neural Networks for Greener and Faster Dependency Parsing

The carbon footprint of natural language processing research has been increasing in recent years due to its reliance on large and inefficient neural network implementations. Distillation is a network compression technique which attempts to impart knowledge from a large model to a smaller one. We use teacher-student distillation to improve the efficiency of the Biaffine dependency parser which obtains state-of-the-art performance with respect to accuracy and parsing speed (Dozat and Manning, 2017). When distilling to 20\% of the original model's trainable parameters, we only observe an average decrease of $\sim$1 point for both UAS and LAS across a number of diverse Universal Dependency treebanks while being 2.30x (1.19x) faster than the baseline model on CPU (GPU) at inference time. We also observe a small increase in performance when compressing to 80\% for some treebanks. Finally, through distillation we attain a parser which is not only faster but also more accurate than the fastest modern parser on the Penn Treebank.

arXiv:2006.00816

Real-Time Face and Landmark Localization for Eyeblink Detection

Pavlovian eyeblink conditioning is a powerful experiment used in the field of neuroscience to measure multiple aspects of how we learn in our daily life. To track the movement of the eyelid during an experiment, researchers have traditionally made use of potentiometers or electromyography. More recently, the use of computer vision and image processing alleviated the need for these techniques but currently employed methods require human intervention and are not fast enough to enable real-time processing. In this work, a face- and landmark-detection algorithm have been carefully combined in order to provide fully automated eyelid tracking, and have further been accelerated to make the first crucial step towards online, closed-loop experiments. Such experiments have not been achieved so far and are expected to offer significant insights in the workings of neurological and psychiatric disorders. Based on an extensive literature search, various different algorithms for face detection and landmark detection have been analyzed and evaluated. Two algorithms were identified as most suitable for eyelid detection: the Histogram-of-Oriented-Gradients (HOG) algorithm for face detection and the Ensemble-of-Regression-Trees (ERT) algorithm for landmark detection. These two algorithms have been accelerated on GPU and CPU, achieving speedups of 1,753$\times$ and 11$\times$, respectively. To demonstrate the usefulness of our eyelid-detection algorithm, a research hypothesis was formed and a well-established neuroscientific experiment was employed: eyeblink detection. Our experimental evaluation reveals an overall application runtime of 0.533 ms per frame, which is 1,101$\times$ faster than the sequential implementation and well within the real-time requirements of eyeblink conditioning in humans, i.e. faster than 500 frames per second.

arXiv:2006.00568

A General-Purpose Dehazing Algorithm based on Local Contrast Enhancement Approaches

Dehazing is in the image processing and computer vision communities, the task of enhancing the image taken in foggy conditions. To better understand this type of algorithm, we present in this document a dehazing method which is suitable for several local contrast adjustment algorithms. We base it on two filters. The first filter is built with a step of normalization with some other statistical tricks while the last represents the local contrast improvement algorithm. Thus, it can work on both CPU and GPU for real-time applications. We hope that our approach will open the door to new ideas in the community. Other advantages of our method are first that it does not need to be trained, then it does not need additional optimization processing. Furthermore, it can be used as a pre-treatment or post-processing step in many vision tasks. In addition, it does not need to convert the problem into a physical interpretation, and finally that it is very fast. This family of defogging algorithms is fairly simple, but it shows promising results compared to state-of-the-art algorithms based not only on a visual assessment but also on objective criteria.