PyTorchKR Papers

최신 논문

총 5,575편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2008.07828

Mastering Large Scale Multi-label Image Recognition with high efficiency overCamera trap images

Camera traps are crucial in biodiversity motivated studies, however dealing with large number of images while annotating these data sets is a tedious and time consuming task. To speed up this process, Machine Learning approaches are a reasonable asset. In this article we are proposing an easy, accessible, light-weight, fast and efficient approach based on our winning submission to the "Hakuna Ma-data - Serengeti Wildlife Identification challenge". Our system achieved an Accuracy of 97% and outperformed the human level performance. We show that, given relatively large data sets, it is effective to look at each image only once with little or no augmentation. By utilizing such a simple, yet effective baseline we were able to avoid over-fitting without extensive regularization techniques and to train a top scoring system on a very limited hardware featuring single GPU (1080Ti) despite the large training set (6.7M images and 6TB).

arXiv:2008.06142공식 코드 · Papers with Code

Landmark detection in Cardiac Magnetic Resonance Imaging Using A Convolutional Neural Network

Purpose: To develop a convolutional neural network (CNN) solution for robust landmark detection in cardiac MR images. Methods: This retrospective study included cine, LGE and T1 mapping scans from two hospitals. The training set included 2,329 patients and 34,019 images. A hold-out test set included 531 patients and 7,723 images. CNN models were developed to detect two mitral valve plane and apical points on long-axis (LAX) images. On short-axis (SAX) images, anterior and posterior RV insertion points and LV center were detected. Model outputs were compared to manual labels by two operators for accuracy with a t-test for statistical significance. The trained model was deployed to MR scanners. Results: For the LAX images, success detection was 99.8% for cine, 99.4% for LGE. For the SAX, success rate was 96.6%, 97.6% and 98.9% for cine, LGE and T1-mapping. The L2 distances between model and manual labels were 2 to 3.5 mm, indicating close agreement between model landmarks to manual labels. No significant differences were found for the anterior RV insertion angle and LV length by the models and operators for all views and imaging sequences. Model inference on MR scanner took 610ms/5.6s on GPU/CPU, respectively, for a typical cardiac cine series. Conclusions: This study developed, validated and deployed a CNN solution for robust landmark detection in both long and short-axis CMR images for cine, LGE and T1 mapping sequences, with the accuracy comparable to the inter-operator variation.

arXiv:2008.06135

Learnability and Robustness of Shallow Neural Networks Learned With a Performance-Driven BP and a Variant PSO For Edge Decision-Making

In many cases, the computing resources are limited without the benefit from GPU, especially in the edge devices of IoT enabled systems. It may not be easy to implement complex AI models in edge devices. The Universal Approximation Theorem states that a shallow neural network (SNN) can represent any nonlinear function. However, how fat is an SNN enough to solve a nonlinear decision-making problem in edge devices? In this paper, we focus on the learnability and robustness of SNNs, obtained by a greedy tight force heuristic algorithm (performance driven BP) and a loose force meta-heuristic algorithm (a variant of PSO). Two groups of experiments are conducted to examine the learnability and the robustness of SNNs with Sigmoid activation, learned/optimised by KPI-PDBPs and KPI-VPSOs, where, KPIs (key performance indicators: error (ERR), accuracy (ACC) and $F_1$ score) are the objectives, driving the searching process. An incremental approach is applied to examine the impact of hidden neuron numbers on the performance of SNNs, learned/optimised by KPI-PDBPs and KPI-VPSOs. From the engineering prospective, all sensors are well justified for a specific task. Hence, all sensor readings should be strongly correlated to the target. Therefore, the structure of an SNN should depend on the dimensions of a problem space. The experimental results show that the number of hidden neurons up to the dimension number of a problem space is enough; the learnability of SNNs, produced by KPI-PDBP, is better than that of SNNs, optimized by KPI-VPSO, regarding the performance and learning time on the training data sets; the robustness of SNNs learned by KPI-PDBPs and KPI-VPSOs depends on the data sets; and comparing with other classic machine learning models, ACC-PDBPs win for almost all tested data sets.

arXiv:2008.06534공식 코드 · Papers with Code

MatryODShka: Real-time 6DoF Video View Synthesis using Multi-Sphere Images

We introduce a method to convert stereo 360{\deg} (omnidirectional stereo) imagery into a layered, multi-sphere image representation for six degree-of-freedom (6DoF) rendering. Stereo 360{\deg} imagery can be captured from multi-camera systems for virtual reality (VR), but lacks motion parallax and correct-in-all-directions disparity cues. Together, these can quickly lead to VR sickness when viewing content. One solution is to try and generate a format suitable for 6DoF rendering, such as by estimating depth. However, this raises questions as to how to handle disoccluded regions in dynamic scenes. Our approach is to simultaneously learn depth and disocclusions via a multi-sphere image representation, which can be rendered with correct 6DoF disparity and motion parallax in VR. This significantly improves comfort for the viewer, and can be inferred and rendered in real time on modern GPU hardware. Together, these move towards making VR video a more comfortable immersive medium.

arXiv:2008.07119

Generative Design by Reinforcement Learning: Enhancing the Diversity of Topology Optimization Designs

Generative design refers to computational design methods that can automatically conduct design exploration under constraints defined by designers. Among many approaches, topology optimization-based generative designs aim to explore diverse topology designs, which cannot be represented by conventional parametric design approaches. Recently, data-driven topology optimization research has started to exploit artificial intelligence, such as deep learning or machine learning, to improve the capability of design exploration. This study proposes a reinforcement learning (RL) based generative design process, with reward functions maximizing the diversity of topology designs. We formulate generative design as a sequential problem of finding optimal design parameter combinations in accordance with a given reference design. Proximal Policy Optimization is used as the learning framework, which is demonstrated in the case study of an automotive wheel design problem. To reduce the heavy computational burden of the wheel topology optimization process required by our RL formulation, we approximate the optimization process with neural networks. With efficient data preprocessing/augmentation and neural architecture, the neural networks achieve a generalized performance and symmetricity-reserving characteristics. We show that RL-based generative design produces a large number of diverse designs within a short inference time by exploiting GPU in a fully automated manner. It is different from the previous approach using CPU which takes much more processing time and involving human intervention.

arXiv:2008.07018

AutoPose: Searching Multi-Scale Branch Aggregation for Pose Estimation

We present AutoPose, a novel neural architecture search(NAS) framework that is capable of automatically discovering multiple parallel branches of cross-scale connections towards accurate and high-resolution 2D human pose estimation. Recently, high-performance hand-crafted convolutional networks for pose estimation show growing demands on multi-scale fusion and high-resolution representations. However, current NAS works exhibit limited flexibility on scale searching, they dominantly adopt simplified search spaces of single-branch architectures. Such simplification limits the fusion of information at different scales and fails to maintain high-resolution representations. The presentedAutoPose framework is able to search for multi-branch scales and network depth, in addition to the cell-level microstructure. Motivated by the search space, a novel bi-level optimization method is presented, where the network-level architecture is searched via reinforcement learning, and the cell-level search is conducted by the gradient-based method. Within 2.5 GPU days, AutoPose is able to find very competitive architectures on the MS COCO dataset, that are also transferable to the MPII dataset. Our code is available at https://github.com/VITA-Group/AutoPose.

arXiv:2008.07008

Monocular Instance Motion Segmentation for Autonomous Driving: KITTI InstanceMotSeg Dataset and Multi-task Baseline

Moving object segmentation is a crucial task for autonomous vehicles as it can be used to segment objects in a class agnostic manner based on their motion cues. It enables the detection of unseen objects during training (e.g., moose or a construction truck) based on their motion and independent of their appearance. Although pixel-wise motion segmentation has been studied in autonomous driving literature, it has been rarely addressed at the instance level, which would help separate connected segments of moving objects leading to better trajectory planning. As the main issue is the lack of large public datasets, we create a new InstanceMotSeg dataset comprising of 12.9K samples improving upon our KITTIMoSeg dataset. In addition to providing instance level annotations, we have added 4 additional classes which is crucial for studying class agnostic motion segmentation. We adapt YOLACT and implement a motion-based class agnostic instance segmentation model which would act as a baseline for the dataset. We also extend it to an efficient multi-task model which additionally provides semantic instance segmentation sharing the encoder. The model then learns separate prototype coefficients within the class agnostic and semantic heads providing two independent paths of object detection for redundant safety. To obtain real-time performance, we study different efficient encoders and obtain 39 fps on a Titan Xp GPU using MobileNetV2 with an improvement of 10% mAP relative to the baseline. Our model improves the previous state of the art motion segmentation method by 3.3%. The dataset and qualitative results video are shared in our website at https://sites.google.com/view/instancemotseg/.

arXiv:2008.06967공식 코드 · Papers with Code

Mesorasi: Architecture Support for Point Cloud Analytics via Delayed-Aggregation

Point cloud analytics is poised to become a key workload on battery-powered embedded and mobile platforms in a wide range of emerging application domains, such as autonomous driving, robotics, and augmented reality, where efficiency is paramount. This paper proposes Mesorasi, an algorithm-architecture co-designed system that simultaneously improves the performance and energy efficiency of point cloud analytics while retaining its accuracy. Our extensive characterizations of state-of-the-art point cloud algorithms show that, while structurally reminiscent of convolutional neural networks (CNNs), point cloud algorithms exhibit inherent compute and memory inefficiencies due to the unique characteristics of point cloud data. We propose delayed-aggregation, a new algorithmic primitive for building efficient point cloud algorithms. Delayed-aggregation hides the performance bottlenecks and reduces the compute and memory redundancies by exploiting the approximately distributive property of key operations in point cloud algorithms. Delayed-aggregation let point cloud algorithms achieve 1.6x speedup and 51.1% energy reduction on a mobile GPU while retaining the accuracy (-0.9% loss to 1.2% gains). To maximize the algorithmic benefits, we propose minor extensions to contemporary CNN accelerators, which can be integrated into a mobile Systems-on-a-Chip (SoC) without modifying other SoC components. With additional hardware support, Mesorasi achieves up to 3.6x speedup.

arXiv:2008.05416공식 코드 · Papers with Code

DXSLAM: A Robust and Efficient Visual SLAM System with Deep Features

A robust and efficient Simultaneous Localization and Mapping (SLAM) system is essential for robot autonomy. For visual SLAM algorithms, though the theoretical framework has been well established for most aspects, feature extraction and association is still empirically designed in most cases, and can be vulnerable in complex environments. This paper shows that feature extraction with deep convolutional neural networks (CNNs) can be seamlessly incorporated into a modern SLAM framework. The proposed SLAM system utilizes a state-of-the-art CNN to detect keypoints in each image frame, and to give not only keypoint descriptors, but also a global descriptor of the whole image. These local and global features are then used by different SLAM modules, resulting in much more robustness against environmental changes and viewpoint changes compared with using hand-crafted features. We also train a visual vocabulary of local features with a Bag of Words (BoW) method. Based on the local features, global features, and the vocabulary, a highly reliable loop closure detection method is built. Experimental results show that all the proposed modules significantly outperforms the baseline, and the full system achieves much lower trajectory errors and much higher correct rates on all evaluated data. Furthermore, by optimizing the CNN with Intel OpenVINO toolkit and utilizing the Fast BoW library, the system benefits greatly from the SIMD (single-instruction-multiple-data) techniques in modern CPUs. The full system can run in real-time without any GPU or other accelerators. The code is public at https://github.com/ivipsourcecode/dxslam.

arXiv:2008.05314공식 코드 · Papers with Code

TF-NAS: Rethinking Three Search Freedoms of Latency-Constrained Differentiable Neural Architecture Search

With the flourish of differentiable neural architecture search (NAS), automatically searching latency-constrained architectures gives a new perspective to reduce human labor and expertise. However, the searched architectures are usually suboptimal in accuracy and may have large jitters around the target latency. In this paper, we rethink three freedoms of differentiable NAS, i.e. operation-level, depth-level and width-level, and propose a novel method, named Three-Freedom NAS (TF-NAS), to achieve both good classification accuracy and precise latency constraint. For the operation-level, we present a bi-sampling search algorithm to moderate the operation collapse. For the depth-level, we introduce a sink-connecting search space to ensure the mutual exclusion between skip and other candidate operations, as well as eliminate the architecture redundancy. For the width-level, we propose an elasticity-scaling strategy that achieves precise latency constraint in a progressively fine-grained manner. Experiments on ImageNet demonstrate the effectiveness of TF-NAS. Particularly, our searched TF-NAS-A obtains 76.9% top-1 accuracy, achieving state-of-the-art results with less latency. The total search time is only 1.8 days on 1 Titan RTX GPU. Code is available at https://github.com/AberHu/TF-NAS.

arXiv:2008.04567

Woodpecker-DL: Accelerating Deep Neural Networks via Hardware-Aware Multifaceted Optimizations

Accelerating deep model training and inference is crucial in practice. Existing deep learning frameworks usually concentrate on optimizing training speed and pay fewer attentions to inference-specific optimizations. Actually, model inference differs from training in terms of computation, e.g. parameters are refreshed each gradient update step during training, but kept invariant during inference. These special characteristics of model inference open new opportunities for its optimization. In this paper, we propose a hardware-aware optimization framework, namely Woodpecker-DL (WPK), to accelerate inference by taking advantage of multiple joint optimizations from the perspectives of graph optimization, automated searches, domain-specific language (DSL) compiler techniques and system-level exploration. In WPK, we investigated two new automated search approaches based on genetic algorithm and reinforcement learning, respectively, to hunt the best operator code configurations targeting specific hardware. A customized DSL compiler is further attached to these search algorithms to generate efficient codes. To create an optimized inference plan, WPK systematically explores high-speed operator implementations from third-party libraries besides our automatically generated codes and singles out the best implementation per operator for use. Extensive experiments demonstrated that on a Tesla P100 GPU, we can achieve the maximum speedup of 5.40 over cuDNN and 1.63 over TVM on individual convolution operators, and run up to 1.18 times faster than TensorRT for end-to-end model inference.

arXiv:2008.03518

Scalable FastMDP for Pre-departure Airspace Reservation and Strategic De-conflict

Pre-departure flight plan scheduling for Urban Air Mobility (UAM) and cargo delivery drones will require on-demand scheduling of large numbers of aircraft. We examine the scalability of an algorithm known as FastMDP which was shown to perform well in deconflicting many dozens of aircraft in a dense airspace environment with terrain. We show that the algorithm can adapted to perform first-come-first-served pre-departure flight plan scheduling where conflict free flight plans are generated on demand. We demonstrate a parallelized implementation of the algorithm on a Graphics Processor Unit (GPU) which we term FastMDP-GPU and show the level of performance and scaling that can be achieved. Our results show that on commodity GPU hardware we can perform flight plan scheduling against 2000-3000 known flight plans and with server-class hardware the performance can be higher. We believe the results show promise for implementing a large scale UAM scheduler capable of performing on-demand flight scheduling that would be suitable for both a centralized or distributed flight planning system

arXiv:2008.03433

GPU-Accelerated Primal Learning for Extremely Fast Large-Scale Classification

One of the most efficient methods to solve L2-regularized primal problems, such as logistic regression and linear support vector machine (SVM) classification, is the widely used trust region Newton algorithm, TRON. While TRON has recently been shown to enjoy substantial speedups on shared-memory multi-core systems, exploiting graphical processing units (GPUs) to speed up the method is significantly more difficult, owing to the highly complex and heavily sequential nature of the algorithm. In this work, we show that using judicious GPU-optimization principles, TRON training time for different losses and feature representations may be drastically reduced. For sparse feature sets, we show that using GPUs to train logistic regression classifiers in LIBLINEAR is up to an order-of-magnitude faster than solely using multithreading. For dense feature sets--which impose far more stringent memory constraints--we show that GPUs substantially reduce the lengthy SVM learning times required for state-of-the-art proteomics analysis, leading to dramatic improvements over recently proposed speedups. Furthermore, we show how GPU speedups may be mixed with multithreading to enable such speedups when the dataset is too large for GPU memory requirements; on a massive dense proteomics dataset of nearly a quarter-billion data instances, these mixed-architecture speedups reduce SVM analysis time from over half a week to less than a single day while using limited GPU memory.

arXiv:2008.03802공식 코드 · Papers with Code

SpeedySpeech: Efficient Neural Speech Synthesis

While recent neural sequence-to-sequence models have greatly improved the quality of speech synthesis, there has not been a system capable of fast training, fast inference and high-quality audio synthesis at the same time. We propose a student-teacher network capable of high-quality faster-than-real-time spectrogram synthesis, with low requirements on computational resources and fast training time. We show that self-attention layers are not necessary for generation of high quality audio. We utilize simple convolutional blocks with residual connections in both student and teacher networks and use only a single attention layer in the teacher model. Coupled with a MelGAN vocoder, our model's voice quality was rated significantly higher than Tacotron 2. Our model can be efficiently trained on a single GPU and can run in real time even on a CPU. We provide both our source code and audio samples in our GitHub repository.

arXiv:2008.03750

Switching Loss for Generalized Nucleus Detection in Histopathology

The accuracy of deep learning methods for two foundational tasks in medical image analysis -- detection and segmentation -- can suffer from class imbalance. We propose a `switching loss' function that adaptively shifts the emphasis between foreground and background classes. While the existing loss functions to address this problem were motivated by the classification task, the switching loss is based on Dice loss, which is better suited for segmentation and detection. Furthermore, to get the most out the training samples, we adapt the loss with each mini-batch, unlike previous proposals that adapt once for the entire training set. A nucleus detector trained using the proposed loss function on a source dataset outperformed those trained using cross-entropy, Dice, or focal losses. Remarkably, without retraining on target datasets, our pre-trained nucleus detector also outperformed existing nucleus detectors that were trained on at least some of the images from the target datasets. To establish a broad utility of the proposed loss, we also confirmed that it led to more accurate ventricle segmentation in MRI as compared to the other loss functions. Our GPU-enabled pre-trained nucleus detection software is also ready to process whole slide images right out-of-the-box and is usably fast.

arXiv:2008.03736공식 코드 · Papers with Code

Fast and Accurate Neural CRF Constituency Parsing

Estimating probability distribution is one of the core issues in the NLP field. However, in both deep learning (DL) and pre-DL eras, unlike the vast applications of linear-chain CRF in sequence labeling tasks, very few works have applied tree-structure CRF to constituency parsing, mainly due to the complexity and inefficiency of the inside-outside algorithm. This work presents a fast and accurate neural CRF constituency parser. The key idea is to batchify the inside algorithm for loss computation by direct large tensor operations on GPU, and meanwhile avoid the outside algorithm for gradient computation via efficient back-propagation. We also propose a simple two-stage bracketing-then-labeling parsing approach to improve efficiency further. To improve the parsing performance, inspired by recent progress in dependency parsing, we introduce a new scoring architecture based on boundary representation and biaffine attention, and a beneficial dropout strategy. Experiments on PTB, CTB5.1, and CTB7 show that our two-stage CRF parser achieves new state-of-the-art performance on both settings of w/o and w/ BERT, and can parse over 1,000 sentences per second. We release our code at https://github.com/yzhangcs/crfpar.

arXiv:2008.03602

Spatial Sharing of GPU for Autotuning DNN models

GPUs are used for training, inference, and tuning the machine learning models. However, Deep Neural Network (DNN) vary widely in their ability to exploit the full power of high-performance GPUs. Spatial sharing of GPU enables multiplexing several DNNs on the GPU and can improve GPU utilization, thus improving throughput and lowering latency. DNN models given just the right amount of GPU resources can still provide low inference latency, just as much as dedicating all of the GPU for their inference task. An approach to improve DNN inference is tuning of the DNN model. Autotuning frameworks find the optimal low-level implementation for a certain target device based on the trained machine learning model, thus reducing the DNN's inference latency and increasing inference throughput. We observe an interdependency between the tuned model and its inference latency. A DNN model tuned with specific GPU resources provides the best inference latency when inferred with close to the same amount of GPU resources. While a model tuned with the maximum amount of the GPU's resources has poorer inference latency once the GPU resources are limited for inference. On the other hand, a model tuned with an appropriate amount of GPU resources still achieves good inference latency across a wide range of GPU resource availability. We explore the causes that impact the tuning of a model at different amounts of GPU resources. We present many techniques to maximize resource utilization and improve tuning performance. We enable controlled spatial sharing of GPU to multiplex several tuning applications on the GPU. We scale the tuning server instances and shard the tuning model across multiple client instances for concurrent tuning of different operators of a model, achieving better GPU multiplexing. With our improvements, we decrease DNN autotuning time by up to 75 percent and increase throughput by a factor of 5.

arXiv:2008.02734공식 코드 · Papers with Code

Exact, Parallelizable Dynamic Time Warping Alignment with Linear Memory

Audio alignment is a fundamental preprocessing step in many MIR pipelines. For two audio clips with M and N frames, respectively, the most popular approach, dynamic time warping (DTW), has O(MN) requirements in both memory and computation, which is prohibitive for frame-level alignments at reasonable rates. To address this, a variety of memory efficient algorithms exist to approximate the optimal alignment under the DTW cost. To our knowledge, however, no exact algorithms exist that are guaranteed to break the quadratic memory barrier. In this work, we present a divide and conquer algorithm that computes the exact globally optimal DTW alignment using O(M+N) memory. Its runtime is still O(MN), trading off memory for a 2x increase in computation. However, the algorithm can be parallelized up to a factor of min(M, N) with the same memory constraints, so it can still run more efficiently than the textbook version with an adequate GPU. We use our algorithm to compute exact alignments on a collection of orchestral music, which we use as ground truth to benchmark the alignment accuracy of several popular approximate alignment schemes at scales that were not previously possible.

arXiv:2008.02002

Fast top-K Cosine Similarity Search through XOR-Friendly Binary Quantization on GPUs

We explore the use of GPU for accelerating large scale nearest neighbor search and we propose a fast vector-quantization-based exhaustive nearest neighbor search algorithm that can achieve high accuracy without any indexing construction specifically designed for cosine similarity. This algorithm uses a novel XOR-friendly binary quantization method to encode floating-point numbers such that high-complexity multiplications can be optimized as low-complexity bitwise operations. Experiments show that, our quantization method takes short preprocessing time, and helps make the search speed of our exhaustive search method much more faster than that of popular approximate nearest neighbor algorithms when high accuracy is needed.

arXiv:2008.01532

A Study on Effects of Implicit and Explicit Language Model Information for DBLSTM-CTC Based Handwriting Recognition

Deep Bidirectional Long Short-Term Memory (D-BLSTM) with a Connectionist Temporal Classification (CTC) output layer has been established as one of the state-of-the-art solutions for handwriting recognition. It is well known that the DBLSTM trained by using a CTC objective function will learn both local character image dependency for character modeling and long-range contextual dependency for implicit language modeling. In this paper, we study the effects of implicit and explicit language model information for DBLSTM-CTC based handwriting recognition by comparing the performance of using or without using an explicit language model in decoding. It is observed that even using one million lines of training sentences to train the DBLSTM, using an explicit language model is still helpful. To deal with such a large-scale training problem, a GPU-based training tool has been developed for CTC training of DBLSTM by using a mini-batch based epochwise Back Propagation Through Time (BPTT) algorithm.