PyTorchKR Papers

최신 논문

총 5,575편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:1912.12607

Towards Unified INT8 Training for Convolutional Neural Network

Recently low-bit (e.g., 8-bit) network quantization has been extensively studied to accelerate the inference. Besides inference, low-bit training with quantized gradients can further bring more considerable acceleration, since the backward process is often computation-intensive. Unfortunately, the inappropriate quantization of backward propagation usually makes the training unstable and even crash. There lacks a successful unified low-bit training framework that can support diverse networks on various tasks. In this paper, we give an attempt to build a unified 8-bit (INT8) training framework for common convolutional neural networks from the aspects of both accuracy and speed. First, we empirically find the four distinctive characteristics of gradients, which provide us insightful clues for gradient quantization. Then, we theoretically give an in-depth analysis of the convergence bound and derive two principles for stable INT8 training. Finally, we propose two universal techniques, including Direction Sensitive Gradient Clipping that reduces the direction deviation of gradients and Deviation Counteractive Learning Rate Scaling that avoids illegal gradient update along the wrong direction. The experiments show that our unified solution promises accurate and efficient INT8 training for a variety of networks and tasks, including MobileNetV2, InceptionV3 and object detection that prior studies have never succeeded. Moreover, it enjoys a strong flexibility to run on off-the-shelf hardware, and reduces the training time by 22% on Pascal GPU without too much optimization effort. We believe that this pioneering study will help lead the community towards a fully unified INT8 training for convolutional neural networks.

arXiv:1912.12055공식 코드 · Papers with Code

nnAudio: An on-the-fly GPU Audio to Spectrogram Conversion Toolbox Using 1D Convolution Neural Networks

Converting time domain waveforms to frequency domain spectrograms is typically considered to be a prepossessing step done before model training. This approach, however, has several drawbacks. First, it takes a lot of hard disk space to store different frequency domain representations. This is especially true during the model development and tuning process, when exploring various types of spectrograms for optimal performance. Second, if another dataset is used, one must process all the audio clips again before the network can be retrained. In this paper, we integrate the time domain to frequency domain conversion as part of the model structure, and propose a neural network based toolbox, nnAudio, which leverages 1D convolutional neural networks to perform time domain to frequency domain conversion during feed-forward. It allows on-the-fly spectrogram generation without the need to store any spectrograms on the disk. This approach also allows back-propagation on the waveforms-to-spectrograms transformation layer, which implies that this transformation process can be made trainable, and hence further optimized by gradient descent. nnAudio reduces the waveforms-to-spectrograms conversion time for 1,770 waveforms (from the MAPS dataset) from $10.64$ seconds with librosa to only $0.001$ seconds for Short-Time Fourier Transform (STFT), $18.3$ seconds to $0.015$ seconds for Mel spectrogram, $103.4$ seconds to $0.258$ for constant-Q transform (CQT), when using GPU on our DGX work station with CPU: Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz Tesla v100 32Gb GPUs. (Only 1 GPU is being used for all the experiments.) We also further optimize the existing CQT algorithm, so that the CQT spectrogram can be obtained without aliasing in a much faster computation time (from $0.258$ seconds to only $0.001$ seconds).

arXiv:1912.11844

Efficient Video Semantic Segmentation with Labels Propagation and Refinement

This paper tackles the problem of real-time semantic segmentation of high definition videos using a hybrid GPU / CPU approach. We propose an Efficient Video Segmentation(EVS) pipeline that combines: (i) On the CPU, a very fast optical flow method, that is used to exploit the temporal aspect of the video and propagate semantic information from one frame to the next. It runs in parallel with the GPU. (ii) On the GPU, two Convolutional Neural Networks: A main segmentation network that is used to predict dense semantic labels from scratch, and a Refiner that is designed to improve predictions from previous frames with the help of a fast Inconsistencies Attention Module (IAM). The latter can identify regions that cannot be propagated accurately. We suggest several operating points depending on the desired frame rate and accuracy. Our pipeline achieves accuracy levels competitive to the existing real-time methods for semantic image segmentation(mIoU above 60%), while achieving much higher frame rates. On the popular Cityscapes dataset with high resolution frames (2048 x 1024), the proposed operating points range from 80 to 1000 Hz on a single GPU and CPU.

arXiv:1912.11350

Atmospheric turbulence removal using convolutional neural network

This paper describes a novel deep learning-based method for mitigating the effects of atmospheric distortion. We have built an end-to-end supervised convolutional neural network (CNN) to reconstruct turbulence-corrupted video sequence. Our framework has been developed on the residual learning concept, where the spatio-temporal distortions are learnt and predicted. Our experiments demonstrate that the proposed method can deblur, remove ripple effect and enhance contrast of the video sequences simultaneously. Our model was trained and tested with both simulated and real distortions. Experimental results of the real distortions show that our method outperforms the existing ones by up to 3.8% in term of the quality of restored images, and it achieves faster speed than the state-of-the-art methods by up to 23 times with GPU implementation.

arXiv:1912.11040

end-to-end training of a large vocabulary end-to-end speech recognition system

In this paper, we present an end-to-end training framework for building state-of-the-art end-to-end speech recognition systems. Our training system utilizes a cluster of Central Processing Units(CPUs) and Graphics Processing Units (GPUs). The entire data reading, large scale data augmentation, neural network parameter updates are all performed "on-the-fly". We use vocal tract length perturbation [1] and an acoustic simulator [2] for data augmentation. The processed features and labels are sent to the GPU cluster. The Horovod allreduce approach is employed to train neural network parameters. We evaluated the effectiveness of our system on the standard Librispeech corpus [3] and the 10,000-hr anonymized Bixby English dataset. Our end-to-end speech recognition system built using this training infrastructure showed a 2.44 % WER on test-clean of the LibriSpeech test set after applying shallow fusion with a Transformer language model (LM). For the proprietary English Bixby open domain test set, we obtained a WER of 7.92 % using a Bidirectional Full Attention (BFA) end-to-end model after applying shallow fusion with an RNN-LM. When the monotonic chunckwise attention (MoCha) based approach is employed for streaming speech recognition, we obtained a WER of 9.95 % on the same Bixby open domain test set.

arXiv:1912.10952공식 코드 · Papers with Code

Progressive DARTS: Bridging the Optimization Gap for NAS in the Wild

With the rapid development of neural architecture search (NAS), researchers found powerful network architectures for a wide range of vision tasks. However, it remains unclear if the searched architecture can transfer across different types of tasks as manually designed ones did. This paper puts forward this problem, referred to as NAS in the wild, which explores the possibility of finding the optimal architecture in a proxy dataset and then deploying it to mostly unseen scenarios. We instantiate this setting using a currently popular algorithm named differentiable architecture search (DARTS), which often suffers unsatisfying performance while being transferred across different tasks. We argue that the accuracy drop originates from the formulation that uses a super-network for search but a sub-network for re-training. The different properties of these stages have resulted in a significant optimization gap, and consequently, the architectural parameters "over-fit" the super-network. To alleviate the gap, we present a progressive method that gradually increases the network depth during the search stage, which leads to the Progressive DARTS (P-DARTS) algorithm. With a reduced search cost (7 hours on a single GPU), P-DARTS achieves improved performance on both the proxy dataset (CIFAR10) and a few target problems (ImageNet classification, COCO detection and three ReID benchmarks). Our code is available at \url{https://github.com/chenxin061/pdarts}.

arXiv:1912.10269공식 코드 · Papers with Code

UWGAN: Underwater GAN for Real-world Underwater Color Restoration and Dehazing

In real-world underwater environment, exploration of seabed resources, underwater archaeology, and underwater fishing rely on a variety of sensors, vision sensor is the most important one due to its high information content, non-intrusive, and passive nature. However, wavelength-dependent light attenuation and back-scattering result in color distortion and haze effect, which degrade the visibility of images. To address this problem, firstly, we proposed an unsupervised generative adversarial network (GAN) for generating realistic underwater images (color distortion and haze effect) from in-air image and depth map pairs based on improved underwater imaging model. Secondly, U-Net, which is trained efficiently using synthetic underwater dataset, is adopted for color restoration and dehazing. Our model directly reconstructs underwater clear images using end-to-end autoencoder networks, while maintaining scene content structural similarity. The results obtained by our method were compared with existing methods qualitatively and quantitatively. Experimental results obtained by the proposed model demonstrate well performance on open real-world underwater datasets, and the processing speed can reach up to 125FPS running on one NVIDIA 1060 GPU. Source code, sample datasets are made publicly available at https://github.com/infrontofme/UWGAN_UIE.

arXiv:1912.09268공식 코드 · Papers with Code

MG-WFBP: Merging Gradients Wisely for Efficient Communication in Distributed Deep Learning

Distributed synchronous stochastic gradient descent has been widely used to train deep neural networks (DNNs) on computer clusters. With the increase of computational power, network communications generally limit the system scalability. Wait-free backpropagation (WFBP) is a popular solution to overlap communications with computations during the training process. In this paper, we observe that many DNNs have a large number of layers with only a small amount of data to be communicated at each layer in distributed training, which could make WFBP inefficient. Based on the fact that merging some short communication tasks into a single one can reduce the overall communication time, we formulate an optimization problem to minimize the training time in pipelining communications and computations. We derive an optimal solution that can be solved efficiently without affecting the training performance. We then apply the solution to propose a distributed training algorithm named merged-gradient WFBP (MG-WFBP) and implement it in two platforms Caffe and PyTorch. Extensive experiments in three GPU clusters are conducted to verify the effectiveness of MG-WFBP. We further exploit trace-based simulations of 4 to 2048 GPUs to explore the potential scaling efficiency of MG-WFBP. Experimental results show that MG-WFBP achieves much better scaling performance than existing methods.

arXiv:1912.08860

Lower Dimensional Kernels for Video Discriminators

This work presents an analysis of the discriminators used in Generative Adversarial Networks (GANs) for Video. We show that unconstrained video discriminator architectures induce a loss surface with high curvature which make optimisation difficult. We also show that this curvature becomes more extreme as the maximal kernel dimension of video discriminators increases. With these observations in hand, we propose a family of efficient Lower-Dimensional Video Discriminators for GANs (LDVD GANs). The proposed family of discriminators improve the performance of video GAN models they are applied to and demonstrate good performance on complex and diverse datasets such as UCF-101. In particular, we show that they can double the performance of Temporal-GANs and provide for state-of-the-art performance on a single GPU.

arXiv:1912.08393공식 코드 · Papers with Code

Salient Object Detection with Purificatory Mechanism and Structural Similarity Loss

By the aid of attention mechanisms to weight the image features adaptively, recent advanced deep learning-based models encourage the predicted results to approximate the ground-truth masks with as large predictable areas as possible, thus achieving the state-of-the-art performance. However, these methods do not pay enough attention to small areas prone to misprediction. In this way, it is still tough to accurately locate salient objects due to the existence of regions with indistinguishable foreground and background and regions with complex or fine structures. To address these problems, we propose a novel convolutional neural network with purificatory mechanism and structural similarity loss. Specifically, in order to better locate preliminary salient objects, we first introduce the promotion attention, which is based on spatial and channel attention mechanisms to promote attention to salient regions. Subsequently, for the purpose of restoring the indistinguishable regions that can be regarded as error-prone regions of one model, we propose the rectification attention, which is learned from the areas of wrong prediction and guide the network to focus on error-prone regions thus rectifying errors. Through these two attentions, we use the Purificatory Mechanism to impose strict weights with different regions of the whole salient objects and purify results from hard-to-distinguish regions, thus accurately predicting the locations and details of salient objects. In addition to paying different attention to these hard-to-distinguish regions, we also consider the structural constraints on complex regions and propose the Structural Similarity Loss. In experiments, the proposed approach outperforms 19 state-of-the-art methods on six datasets with a notable margin at over 27FPS on a single NVIDIA 1080Ti GPU.

논문 ID: document/8091247

End-to-End Deep Learning of Optimization Heuristics

Accurate automatic optimization heuristics are necessary for dealing with the complexity and diversity of modern hardware and software. Machine learning is a proven technique for learning such heuristics, but its success is bound by the quality of the features used. These features must be hand crafted by developers through a combination of expert domain knowledge and trial and error. This makes the quality of the final model directly dependent on the skill and available time of the system architect. Our work introduces a better way for building heuristics. We develop a deep neural network that learns heuristics over raw code, entirely without using code features. The neural network simultaneously constructs appropriate representations of the code and learns how best to optimize, removing the need for manual feature creation. Further, we show that our neural nets can transfer learning from one optimization problem to another, improving the accuracy of new models, without the help of human experts. We compare the effectiveness of our automatically generated heuristics against ones with features hand-picked by experts. We examine two challenging tasks:predicting optimal mapping for heterogeneous parallelism and GPU thread coarsening factors. In 89% of the cases, the quality of our fully automatic heuristics matches or surpasses that of state-of-the-art predictive models using hand-crafted features, providing on average 14% and 12% more performance with no human effort expended on designing features.

arXiv:1912.07906

Deep SCNN-based Real-time Object Detection for Self-driving Vehicles Using LiDAR Temporal Data

Real-time accurate detection of three-dimensional (3D) objects is a fundamental necessity for self-driving vehicles. Most existing computer vision approaches are based on convolutional neural networks (CNNs). Although the CNN-based approaches can achieve high detection accuracy, their high energy consumption is a severe drawback. To resolve this problem, novel energy efficient approaches should be explored. Spiking neural network (SNN) is a promising candidate because it has orders-of-magnitude lower energy consumption than CNN. Unfortunately, the studying of SNN has been limited in small networks only. The application of SNN for large 3D object detection networks has remain largely open. In this paper, we integrate spiking convolutional neural network (SCNN) with temporal coding into the YOLOv2 architecture for real-time object detection. To take the advantage of spiking signals, we develop a novel data preprocessing layer that translates 3D point-cloud data into spike time data. We propose an analog circuit to implement the non-leaky integrate and fire neuron used in our SCNN, from which the energy consumption of each spike is estimated. Moreover, we present a method to calculate the network sparsity and the energy consumption of the overall network. Extensive experiments have been conducted based on the KITTI dataset, which show that the proposed network can reach competitive detection accuracy as existing approaches, yet with much lower average energy consumption. If implemented in dedicated hardware, our network could have a mean sparsity of 56.24% and extremely low total energy consumption of 0.247mJ only. Implemented in NVIDIA GTX 1080i GPU, we can achieve 35.7 fps frame rate, high enough for real-time object detection.

arXiv:1912.06798공식 코드 · Papers with Code

Cross-Batch Memory for Embedding Learning

Mining informative negative instances are of central importance to deep metric learning (DML), however this task is intrinsically limited by mini-batch training, where only a mini-batch of instances is accessible at each iteration. In this paper, we identify a "slow drift" phenomena by observing that the embedding features drift exceptionally slow even as the model parameters are updating throughout the training process. This suggests that the features of instances computed at preceding iterations can be used to considerably approximate their features extracted by the current model. We propose a cross-batch memory (XBM) mechanism that memorizes the embeddings of past iterations, allowing the model to collect sufficient hard negative pairs across multiple mini-batches - even over the whole dataset. Our XBM can be directly integrated into a general pair-based DML framework, where the XBM augmented DML can boost performance considerably. In particular, without bells and whistles, a simple contrastive loss with our XBM can have large R@1 improvements of 12%-22.5% on three large-scale image retrieval datasets, surpassing the most sophisticated state-of-the-art methods, by a large margin. Our XBM is conceptually simple, easy to implement - using several lines of codes, and is memory efficient - with a negligible 0.2 GB extra GPU memory. Code is available at: https://github.com/MalongTech/research-xbm.

arXiv:1912.06378공식 코드 · Papers with Code

Cascade Cost Volume for High-Resolution Multi-View Stereo and Stereo Matching

The deep multi-view stereo (MVS) and stereo matching approaches generally construct 3D cost volumes to regularize and regress the output depth or disparity. These methods are limited when high-resolution outputs are needed since the memory and time costs grow cubically as the volume resolution increases. In this paper, we propose a both memory and time efficient cost volume formulation that is complementary to existing multi-view stereo and stereo matching approaches based on 3D cost volumes. First, the proposed cost volume is built upon a standard feature pyramid encoding geometry and context at gradually finer scales. Then, we can narrow the depth (or disparity) range of each stage by the depth (or disparity) map from the previous stage. With gradually higher cost volume resolution and adaptive adjustment of depth (or disparity) intervals, the output is recovered in a coarser to fine manner. We apply the cascade cost volume to the representative MVS-Net, and obtain a 23.1% improvement on DTU benchmark (1st place), with 50.6% and 74.2% reduction in GPU memory and run-time. It is also the state-of-the-art learning-based method on Tanks and Temples benchmark. The statistics of accuracy, run-time and GPU memory on other representative stereo CNNs also validate the effectiveness of our proposed method.

arXiv:1912.06322

Queueing Analysis of GPU-Based Inference Servers with Dynamic Batching: A Closed-Form Characterization

GPU-accelerated computing is a key technology to realize high-speed inference servers using deep neural networks (DNNs). An important characteristic of GPU-based inference is that the computational efficiency, in terms of the processing speed and energy consumption, drastically increases by processing multiple jobs together in a batch. In this paper, we formulate GPU-based inference servers as a batch service queueing model with batch-size dependent processing times. We first show that the energy efficiency of the server monotonically increases with the arrival rate of inference jobs, which suggests that it is energy-efficient to operate the inference server under a utilization level as high as possible within a latency requirement of inference jobs. We then derive a closed-form upper bound for the mean latency, which provides a simple characterization of the latency performance. Through simulation and numerical experiments, we show that the exact value of the mean latency is well approximated by this upper bound. We further compare this upper bound with the latency curve measured in real implementation of GPU-based inference servers and we show that the real performance curve is well explained by the derived simple formula.

arXiv:1912.06218공식 코드 · Papers with Code

YOLACT++: Better Real-time Instance Segmentation

We present a simple, fully-convolutional model for real-time (>30 fps) instance segmentation that achieves competitive results on MS COCO evaluated on a single Titan Xp, which is significantly faster than any previous state-of-the-art approach. Moreover, we obtain this result after training on only one GPU. We accomplish this by breaking instance segmentation into two parallel subtasks: (1) generating a set of prototype masks and (2) predicting per-instance mask coefficients. Then we produce instance masks by linearly combining the prototypes with the mask coefficients. We find that because this process doesn't depend on repooling, this approach produces very high-quality masks and exhibits temporal stability for free. Furthermore, we analyze the emergent behavior of our prototypes and show they learn to localize instances on their own in a translation variant manner, despite being fully-convolutional. We also propose Fast NMS, a drop-in 12 ms faster replacement for standard NMS that only has a marginal performance penalty. Finally, by incorporating deformable convolutions into the backbone network, optimizing the prediction head with better anchor scales and aspect ratios, and adding a novel fast mask re-scoring branch, our YOLACT++ model can achieve 34.1 mAP on MS COCO at 33.5 fps, which is fairly close to the state-of-the-art approaches while still running at real-time.

arXiv:1912.06185

Learning Effective Visual Relationship Detector on 1 GPU

We present our winning solution to the Open Images 2019 Visual Relationship challenge. This is the largest challenge of its kind to date with nearly 9 million training images. Challenge task consists of detecting objects and identifying relationships between them in complex scenes. Our solution has three stages, first object detection model is fine-tuned for the challenge classes using a novel weight transfer approach. Then, spatio-semantic and visual relationship models are trained on candidate object pairs. Finally, features and model predictions are combined to generate the final relationship prediction. Throughout the challenge we focused on minimizing the hardware requirements of our architecture. Specifically, our weight transfer approach enables much faster optimization, allowing the entire architecture to be trained on a single GPU in under two days. In addition to efficient optimization, our approach also achieves superior accuracy winning first place out of over 200 teams, and outperforming the second place team by over $5\%$ on the held-out private leaderboard.

arXiv:1912.04441

HR-SAR-Net: A Deep Neural Network for Urban Scene Segmentation from High-Resolution SAR Data

Synthetic aperture radar (SAR) data is becoming increasingly available to a wide range of users through commercial service providers with resolutions reaching 0.5m/px. Segmenting SAR data still requires skilled personnel, limiting the potential for large-scale use. We show that it is possible to automatically and reliably perform urban scene segmentation from next-gen resolution SAR data (0.15m/px) using deep neural networks (DNNs), achieving a pixel accuracy of 95.19% and a mean IoU of 74.67% with data collected over a region of merely 2.2km${}^2$. The presented DNN is not only effective, but is very small with only 63k parameters and computationally simple enough to achieve a throughput of around 500Mpx/s using a single GPU. We further identify that additional SAR receive antennas and data from multiple flights massively improve the segmentation accuracy. We describe a procedure for generating a high-quality segmentation ground truth from multiple inaccurate building and road annotations, which has been crucial to achieving these segmentation results.

arXiv:1912.05234

Array Languages Make Neural Networks Fast

Modern machine learning frameworks are complex: they are typically organised in multiple layers each of which is written in a different language and they depend on a number of external libraries, but at their core they mainly consist of tensor operations. As array-oriented languages provide perfect abstractions to implement tensor operations, we consider a minimalistic machine learning framework that is shallowly embedded in an array-oriented language and we study its productivity and performance. We do this by implementing a state of the art Convolutional Neural Network (CNN) and compare it against implementations in TensorFlow and PyTorch --- two state of the art industrial-strength frameworks. It turns out that our implementation is 2 and 3 times faster, even after fine-tuning the TensorFlow and PyTorch to our hardware --- a 64-core GPU-accelerated machine. The size of all three CNN specifications is the same, about 150 lines of code. Our mini framework is 150 lines of highly reusable hardware-agnostic code that does not depend on external libraries. The compiler for a host array language automatically generates parallel code for a chosen architecture. The key to such a balance between performance and portability lies in the design of the array language; in particular, the ability to express rank-polymorphic operations concisely, yet being able to do optimisations across them. This design builds on very few assumptions, and it is readily transferable to other contexts offering a clean approach to high-performance machine learning.

arXiv:1912.05137공식 코드 · Papers with Code

Is Feature Diversity Necessary in Neural Network Initialization?

Standard practice in training neural networks involves initializing the weights in an independent fashion. The results of recent work suggest that feature "diversity" at initialization plays an important role in training the network. However, other initialization schemes with reduced feature diversity have also been shown to be viable. In this work, we conduct a series of experiments aimed at elucidating the importance of feature diversity at initialization. We show that a complete lack of diversity is harmful to training, but its effects can be counteracted by a relatively small addition of noise - even the noise in standard non-deterministic GPU computations is sufficient. Furthermore, we construct a deep convolutional network with identical features at initialization and almost all of the weights initialized at 0 that can be trained to reach accuracy matching its standard-initialized counterpart.