PyTorchKR Papers

최신 논문

총 5,576편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2103.15386공식 코드 · Papers with Code

Large-Scale Approximate k-NN Graph Construction on GPU

k-nearest neighbor graph is a key data structure in many disciplines such as manifold learning, machine learning and information retrieval, etc. NN-Descent was proposed as an effective solution for the graph construction problem. However, it cannot be directly transplanted to GPU due to the intensive memory accesses required in the approach. In this paper, NN-Descent has been redesigned to adapt to the GPU architecture. In particular, the number of memory accesses has been reduced significantly. The redesign fully exploits the parallelism of the GPU hardware. In the meantime, the genericness as well as the simplicity of NN-Descent are well-preserved. In addition, a simple but effective k-NN graph merge approach is presented. It allows two graphs to be merged efficiently on GPUs. More importantly, it makes the construction of high-quality k-NN graphs for out-of-GPU-memory datasets tractable. The results show that our approach is 100-250x faster than single-thread NN-Descent and is 2.5-5x faster than existing GPU-based approaches.

arXiv:2103.14529공식 코드 · Papers with Code

Real-Time and Accurate Object Detection in Compressed Video by Long Short-term Feature Aggregation

Video object detection is a fundamental problem in computer vision and has a wide spectrum of applications. Based on deep networks, video object detection is actively studied for pushing the limits of detection speed and accuracy. To reduce the computation cost, we sparsely sample key frames in video and treat the rest frames are non-key frames; a large and deep network is used to extract features for key frames and a tiny network is used for non-key frames. To enhance the features of non-key frames, we propose a novel short-term feature aggregation method to propagate the rich information in key frame features to non-key frame features in a fast way. The fast feature aggregation is enabled by the freely available motion cues in compressed videos. Further, key frame features are also aggregated based on optical flow. The propagated deep features are then integrated with the directly extracted features for object detection. The feature extraction and feature integration parameters are optimized in an end-to-end manner. The proposed video object detection network is evaluated on the large-scale ImageNet VID benchmark and achieves 77.2\% mAP, which is on-par with state-of-the-art accuracy, at the speed of 30 FPS using a Titan X GPU. The source codes are available at \url{https://github.com/hustvl/LSFA}.

arXiv:2103.14409

LS-CAT: A Large-Scale CUDA AutoTuning Dataset

The effectiveness of Machine Learning (ML) methods depend on access to large suitable datasets. In this article, we present how we build the LS-CAT (Large-Scale CUDA AutoTuning) dataset sourced from GitHub for the purpose of training NLP-based ML models. Our dataset includes 19 683 CUDA kernels focused on linear algebra. In addition to the CUDA codes, our LS-CAT dataset contains 5 028 536 associated runtimes, with different combinations of kernels, block sizes and matrix sizes. The runtime are GPU benchmarks on both Nvidia GTX 980 and Nvidia T4 systems. This information creates a foundation upon which NLP-based models can find correlations between source-code features and optimal choice of thread block sizes. There are several results that can be drawn out of our LS-CAT database. E.g., our experimental results show that an optimal choice in thread block size can gain an average of 6% for the average case. We thus also analyze how much performance increase can be achieved in general, finding that in 10% of the cases more than 20% performance increase can be achieved by using the optimal block. A description of current and future work is also included.

arXiv:2103.14137공식 코드 · Papers with Code

Optimized Coverage Planning for UV Surface Disinfection

UV radiation has been used as a disinfection strategy to deactivate a wide range of pathogens, but existing irradiation strategies do not ensure sufficient exposure of all environmental surfaces and/or require long disinfection times. We present a near-optimal coverage planner for mobile UV disinfection robots. The formulation optimizes the irradiation time efficiency, while ensuring that a sufficient dosage of radiation is received by each surface. The trajectory and dosage plan are optimized taking collision and light occlusion constraints into account. We propose a two-stage scheme to approximate the solution of the induced NP-hard optimization, and, for efficiency, perform key irradiance and occlusion calculations on a GPU. Empirical results show that our technique achieves more coverage for the same exposure time as strategies for existing UV robots, can be used to compare UV robot designs, and produces near-optimal plans. This is an extended version of the paper originally contributed to ICRA2021.

arXiv:2103.14645공식 코드 · Papers with Code

Baking Neural Radiance Fields for Real-Time View Synthesis

Neural volumetric representations such as Neural Radiance Fields (NeRF) have emerged as a compelling technique for learning to represent 3D scenes from images with the goal of rendering photorealistic images of the scene from unobserved viewpoints. However, NeRF's computational requirements are prohibitive for real-time applications: rendering views from a trained NeRF requires querying a multilayer perceptron (MLP) hundreds of times per ray. We present a method to train a NeRF, then precompute and store (i.e. "bake") it as a novel representation called a Sparse Neural Radiance Grid (SNeRG) that enables real-time rendering on commodity hardware. To achieve this, we introduce 1) a reformulation of NeRF's architecture, and 2) a sparse voxel grid representation with learned feature vectors. The resulting scene representation retains NeRF's ability to render fine geometric details and view-dependent appearance, is compact (averaging less than 90 MB per scene), and can be rendered in real-time (higher than 30 frames per second on a laptop GPU). Actual screen captures are shown in our video.

arXiv:2103.13799

Bertinho: Galician BERT Representations

This paper presents a monolingual BERT model for Galician. We follow the recent trend that shows that it is feasible to build robust monolingual BERT models even for relatively low-resource languages, while performing better than the well-known official multilingual BERT (mBERT). More particularly, we release two monolingual Galician BERT models, built using 6 and 12 transformer layers, respectively; trained with limited resources (~45 million tokens on a single GPU of 24GB). We then provide an exhaustive evaluation on a number of tasks such as POS-tagging, dependency parsing and named entity recognition. For this purpose, all these tasks are cast in a pure sequence labeling setup in order to run BERT without the need to include any additional layers on top of it (we only use an output classification layer to map the contextualized representations into the predicted label). The experiments show that our models, especially the 12-layer one, outperform the results of mBERT in most tasks.

arXiv:2103.13262공식 코드 · Papers with Code

FastMoE: A Fast Mixture-of-Expert Training System

Mixture-of-Expert (MoE) presents a strong potential in enlarging the size of language model to trillions of parameters. However, training trillion-scale MoE requires algorithm and system co-design for a well-tuned high performance distributed training system. Unfortunately, the only existing platform that meets the requirements strongly depends on Google's hardware (TPU) and software (Mesh Tensorflow) stack, and is not open and available to the public, especially GPU and PyTorch communities. In this paper, we present FastMoE, a distributed MoE training system based on PyTorch with common accelerators. The system provides a hierarchical interface for both flexible model design and easy adaption to different applications, such as Transformer-XL and Megatron-LM. Different from direct implementation of MoE models using PyTorch, the training speed is highly optimized in FastMoE by sophisticated high-performance acceleration skills. The system supports placing different experts on multiple GPUs across multiple nodes, enabling enlarging the number of experts linearly against the number of GPUs. The source of FastMoE is available at https://github.com/laekov/fastmoe under Apache-2 license.

arXiv:2103.13044

MSCFNet: A Lightweight Network With Multi-Scale Context Fusion for Real-Time Semantic Segmentation

In recent years, how to strike a good trade-off between accuracy and inference speed has become the core issue for real-time semantic segmentation applications, which plays a vital role in real-world scenarios such as autonomous driving systems and drones. In this study, we devise a novel lightweight network using a multi-scale context fusion (MSCFNet) scheme, which explores an asymmetric encoder-decoder architecture to dispose this problem. More specifically, the encoder adopts some developed efficient asymmetric residual (EAR) modules, which are composed of factorization depth-wise convolution and dilation convolution. Meanwhile, instead of complicated computation, simple deconvolution is applied in the decoder to further reduce the amount of parameters while still maintaining high segmentation accuracy. Also, MSCFNet has branches with efficient attention modules from different stages of the network to well capture multi-scale contextual information. Then we combine them before the final classification to enhance the expression of the features and improve the segmentation efficiency. Comprehensive experiments on challenging datasets have demonstrated that the proposed MSCFNet, which contains only 1.15M parameters, achieves 71.9\% Mean IoU on the Cityscapes testing dataset and can run at over 50 FPS on a single Titan XP GPU configuration.

arXiv:2103.12212

CFPNet: Channel-wise Feature Pyramid for Real-Time Semantic Segmentation

Real-time semantic segmentation is playing a more important role in computer vision, due to the growing demand for mobile devices and autonomous driving. Therefore, it is very important to achieve a good trade-off among performance, model size and inference speed. In this paper, we propose a Channel-wise Feature Pyramid (CFP) module to balance those factors. Based on the CFP module, we built CFPNet for real-time semantic segmentation which applied a series of dilated convolution channels to extract effective features. Experiments on Cityscapes and CamVid datasets show that the proposed CFPNet achieves an effective combination of those factors. For the Cityscapes test dataset, CFPNet achieves 70.1% class-wise mIoU with only 0.55 million parameters and 2.5 MB memory. The inference speed can reach 30 FPS on a single RTX 2080Ti GPU with a 1024x2048-pixel image.

arXiv:2103.11568공식 코드 · Papers with Code

Cluster Contrast for Unsupervised Person Re-Identification

State-of-the-art unsupervised re-ID methods train the neural networks using a memory-based non-parametric softmax loss. Instance feature vectors stored in memory are assigned pseudo-labels by clustering and updated at instance level. However, the varying cluster sizes leads to inconsistency in the updating progress of each cluster. To solve this problem, we present Cluster Contrast which stores feature vectors and computes contrast loss at the cluster level. Our approach employs a unique cluster representation to describe each cluster, resulting in a cluster-level memory dictionary. In this way, the consistency of clustering can be effectively maintained throughout the pipline and the GPU memory consumption can be significantly reduced. Thus, our method can solve the problem of cluster inconsistency and be applicable to larger data sets. In addition, we adopt different clustering algorithms to demonstrate the robustness and generalization of our framework. The application of Cluster Contrast to a standard unsupervised re-ID pipeline achieves considerable improvements of 9.9%, 8.3%, 12.1% compared to state-of-the-art purely unsupervised re-ID methods and 5.5%, 4.8%, 4.4% mAP compared to the state-of-the-art unsupervised domain adaptation re-ID methods on the Market, Duke, and MSMT17 datasets. Code is available at https://github.com/alibaba/cluster-contrast.

arXiv:2103.09944

IRLI: Iterative Re-partitioning for Learning to Index

Neural models have transformed the fundamental information retrieval problem of mapping a query to a giant set of items. However, the need for efficient and low latency inference forces the community to reconsider efficient approximate near-neighbor search in the item space. To this end, learning to index is gaining much interest in recent times. Methods have to trade between obtaining high accuracy while maintaining load balance and scalability in distributed settings. We propose a novel approach called IRLI (pronounced `early'), which iteratively partitions the items by learning the relevant buckets directly from the query-item relevance data. Furthermore, IRLI employs a superior power-of-$k$-choices based load balancing strategy. We mathematically show that IRLI retrieves the correct item with high probability under very natural assumptions and provides superior load balancing. IRLI surpasses the best baseline's precision on multi-label classification while being $5x$ faster on inference. For near-neighbor search tasks, the same method outperforms the state-of-the-art Learned Hashing approach NeuralLSH by requiring only ~ {1/6}^th of the candidates for the same recall. IRLI is both data and model parallel, making it ideal for distributed GPU implementation. We demonstrate this advantage by indexing 100 million dense vectors and surpassing the popular FAISS library by >10% on recall.

arXiv:2103.10380

FastNeRF: High-Fidelity Neural Rendering at 200FPS

Recent work on Neural Radiance Fields (NeRF) showed how neural networks can be used to encode complex 3D environments that can be rendered photorealistically from novel viewpoints. Rendering these images is very computationally demanding and recent improvements are still a long way from enabling interactive rates, even on high-end hardware. Motivated by scenarios on mobile and mixed reality devices, we propose FastNeRF, the first NeRF-based system capable of rendering high fidelity photorealistic images at 200Hz on a high-end consumer GPU. The core of our method is a graphics-inspired factorization that allows for (i) compactly caching a deep radiance map at each position in space, (ii) efficiently querying that map using ray directions to estimate the pixel values in the rendered image. Extensive experiments show that the proposed method is 3000 times faster than the original NeRF algorithm and at least an order of magnitude faster than existing work on accelerating NeRF, while maintaining visual quality and extensibility.

arXiv:2103.10148공식 코드 · Papers with Code

Sequential End-to-end Network for Efficient Person Search

Person search aims at jointly solving Person Detection and Person Re-identification (re-ID). Existing works have designed end-to-end networks based on Faster R-CNN. However, due to the parallel structure of Faster R-CNN, the extracted features come from the low-quality proposals generated by the Region Proposal Network, rather than the detected high-quality bounding boxes. Person search is a fine-grained task and such inferior features will significantly reduce re-ID performance. To address this issue, we propose a Sequential End-to-end Network (SeqNet) to extract superior features. In SeqNet, detection and re-ID are considered as a progressive process and tackled with two sub-networks sequentially. In addition, we design a robust Context Bipartite Graph Matching (CBGM) algorithm to effectively employ context information as an important complementary cue for person matching. Extensive experiments on two widely used person search benchmarks, CUHK-SYSU and PRW, have shown that our method achieves state-of-the-art results. Also, our model runs at 11.5 fps on a single GPU and can be integrated into the existing end-to-end framework easily.

arXiv:2103.09512공식 코드 · Papers with Code

CNN Model & Tuning for Global Road Damage Detection

This paper provides a report on our solution including model selection, tuning strategy and results obtained for Global Road Damage Detection Challenge. This Big Data Cup Challenge was held as a part of IEEE International Conference on Big Data 2020. We assess single and multi-stage network architectures for object detection and provide a benchmark using popular state-of-the-art open-source PyTorch frameworks like Detectron2 and Yolov5. Data preparation for provided Road Damage training dataset, captured using smartphone camera from Czech, India and Japan is discussed. We studied the effect of training on a per country basis with respect to a single generalizable model. We briefly describe the tuning strategy for the experiments conducted on two-stage Faster R-CNN with Deep Residual Network (Resnet) and Feature Pyramid Network (FPN) backbone. Additionally, we compare this to a one-stage Yolov5 model with Cross Stage Partial Network (CSPNet) backbone. We show a mean F1 score of 0.542 on Test2 and 0.536 on Test1 datasets using a multi-stage Faster R-CNN model, with Resnet-50 and Resnet-101 backbones respectively. This shows the generalizability of the Resnet-50 model when compared to its more complex counterparts. Experiments were conducted using Google Colab having K80 and a Linux PC with 1080Ti, NVIDIA consumer grade GPU. A PyTorch based Detectron2 code to preprocess, train, test and submit the Avg F1 score to is made available at https://github.com/vishwakarmarhl/rdd2020

arXiv:2103.09422공식 코드 · Papers with Code

YOLOStereo3D: A Step Back to 2D for Efficient Stereo 3D Detection

Object detection in 3D with stereo cameras is an important problem in computer vision, and is particularly crucial in low-cost autonomous mobile robots without LiDARs. Nowadays, most of the best-performing frameworks for stereo 3D object detection are based on dense depth reconstruction from disparity estimation, making them extremely computationally expensive. To enable real-world deployments of vision detection with binocular images, we take a step back to gain insights from 2D image-based detection frameworks and enhance them with stereo features. We incorporate knowledge and the inference structure from real-time one-stage 2D/3D object detector and introduce a light-weight stereo matching module. Our proposed framework, YOLOStereo3D, is trained on one single GPU and runs at more than ten fps. It demonstrates performance comparable to state-of-the-art stereo 3D detection frameworks without usage of LiDAR data. The code will be published in https://github.com/Owen-Liuyuxuan/visualDet3D.

arXiv:2103.09404공식 코드 · Papers with Code

Collapsible Linear Blocks for Super-Efficient Super Resolution

With the advent of smart devices that support 4K and 8K resolution, Single Image Super Resolution (SISR) has become an important computer vision problem. However, most super resolution deep networks are computationally very expensive. In this paper, we propose Super-Efficient Super Resolution (SESR) networks that establish a new state-of-the-art for efficient super resolution. Our approach is based on linear overparameterization of CNNs and creates an efficient model architecture for SISR. With theoretical analysis, we uncover the limitations of existing overparameterization methods and show how the proposed method alleviates them. Detailed experiments across six benchmark datasets demonstrate that SESR achieves similar or better image quality than state-of-the-art models while requiring 2x to 330x fewer Multiply-Accumulate (MAC) operations. As a result, SESR can be used on constrained hardware to perform x2 (1080p to 4K) and x4 (1080p to 8K) SISR. Towards this, we estimate hardware performance numbers for a commercial Arm mobile-Neural Processing Unit (NPU) for 1080p to 4K (x2) and 1080p to 8K (x4) SISR. Our results highlight the challenges faced by super resolution on AI accelerators and demonstrate that SESR is significantly faster (e.g., 6x-8x higher FPS) than existing models on mobile-NPU. Finally, SESR outperforms prior models by 1.5x-2x in latency on Arm CPU and GPU when deployed on a real mobile device. The code for this work is available at https://github.com/ARM-software/sesr.

arXiv:2103.08640공식 코드 · Papers with Code

UPANets: Learning from the Universal Pixel Attention Networks

Among image classification, skip and densely-connection-based networks have dominated most leaderboards. Recently, from the successful development of multi-head attention in natural language processing, it is sure that now is a time of either using a Transformer-like model or hybrid CNNs with attention. However, the former need a tremendous resource to train, and the latter is in the perfect balance in this direction. In this work, to make CNNs handle global and local information, we proposed UPANets, which equips channel-wise attention with a hybrid skip-densely-connection structure. Also, the extreme-connection structure makes UPANets robust with a smoother loss landscape. In experiments, UPANets surpassed most well-known and widely-used SOTAs with an accuracy of 96.47% in Cifar-10, 80.29% in Cifar-100, and 67.67% in Tiny Imagenet. Most importantly, these performances have high parameters efficiency and only trained in one customer-based GPU. We share implementing code of UPANets in https://github.com/hanktseng131415go/UPANets.

arXiv:2103.08802

Parareal Neural Networks Emulating a Parallel-in-time Algorithm

As deep neural networks (DNNs) become deeper, the training time increases. In this perspective, multi-GPU parallel computing has become a key tool in accelerating the training of DNNs. In this paper, we introduce a novel methodology to construct a parallel neural network that can utilize multiple GPUs simultaneously from a given DNN. We observe that layers of DNN can be interpreted as the time step of a time-dependent problem and can be parallelized by emulating a parallel-in-time algorithm called parareal. The parareal algorithm consists of fine structures which can be implemented in parallel and a coarse structure which gives suitable approximations to the fine structures. By emulating it, the layers of DNN are torn to form a parallel structure which is connected using a suitable coarse network. We report accelerated and accuracy-preserved results of the proposed methodology applied to VGG-16 and ResNet-1001 on several datasets.

arXiv:2103.07992

Deep Tiling: Texture Tile Synthesis Using a Deep Learning Approach

Texturing is a fundamental process in computer graphics. Texture is leveraged to enhance the visualization outcome for a 3D scene. In many cases a texture image cannot cover a large 3D model surface because of its small resolution. Conventional techniques like repeating, mirror repeating or clamp to edge do not yield visually acceptable results. Deep learning based texture synthesis has proven to be very effective in such cases. All deep texture synthesis methods trying to create larger resolution textures are limited in terms of GPU memory resources. In this paper, we propose a novel approach to example-based texture synthesis by using a robust deep learning process for creating tiles of arbitrary resolutions that resemble the structural components of an input texture. In this manner, our method is firstly much less memory limited owing to the fact that a new texture tile of small size is synthesized and merged with the original texture and secondly can easily produce missing parts of a large texture.

arXiv:2103.08325

HSCoNAS: Hardware-Software Co-Design of Efficient DNNs via Neural Architecture Search

In this paper, we present a novel multi-objective hardware-aware neural architecture search (NAS) framework, namely HSCoNAS, to automate the design of deep neural networks (DNNs) with high accuracy but low latency upon target hardware. To accomplish this goal, we first propose an effective hardware performance modeling method to approximate the runtime latency of DNNs on target hardware, which will be integrated into HSCoNAS to avoid the tedious on-device measurements. Besides, we propose two novel techniques, i.e., dynamic channel scaling to maximize the accuracy under the specified latency and progressive space shrinking to refine the search space towards target hardware as well as alleviate the search overheads. These two techniques jointly work to allow HSCoNAS to perform fine-grained and efficient explorations. Finally, an evolutionary algorithm (EA) is incorporated to conduct the architecture search. Extensive experiments on ImageNet are conducted upon diverse target hardware, i.e., GPU, CPU, and edge device to demonstrate the superiority of HSCoNAS over recent state-of-the-art approaches.