Deep Neural Networks (DNNs) have transformed the field of machine learning and are widely deployed in many applications involving image, video, speech and natural language processing. The increasing compute demands of DNNs have been widely addressed through Graphics Processing Units (GPUs) and specialized accelerators. However, as model sizes grow, these von Neumann architectures require very high memory bandwidth to keep the processing elements utilized as a majority of the data resides in the main memory. Processing in memory has been proposed as a promising solution for the memory wall bottleneck for ML workloads. In this work, we propose a new DRAM-based processing-in-memory (PIM) multiplication primitive coupled with intra-bank accumulation to accelerate matrix vector operations in ML workloads. The proposed multiplication primitive adds < 1% area overhead and does not require any change in the DRAM peripherals. Therefore, the proposed multiplication can be easily adopted in commodity DRAM chips. Subsequently, we design a DRAM-based PIM architecture, data mapping scheme and dataflow for executing DNNs within DRAM. System evaluations performed on networks like AlexNet, VGG16 and ResNet18 show that the proposed architecture, mapping, and data flow can provide up to 19.5x speedup over an NVIDIA Titan Xp GPU highlighting the need to overcome the memory bottleneck in future generations of DNN hardware.
In this paper, we explore the spatial redundancy in video recognition with the aim to improve the computational efficiency. It is observed that the most informative region in each frame of a video is usually a small image patch, which shifts smoothly across frames. Therefore, we model the patch localization problem as a sequential decision task, and propose a reinforcement learning based approach for efficient spatially adaptive video recognition (AdaFocus). In specific, a light-weighted ConvNet is first adopted to quickly process the full video sequence, whose features are used by a recurrent policy network to localize the most task-relevant regions. Then the selected patches are inferred by a high-capacity network for the final prediction. During offline inference, once the informative patch sequence has been generated, the bulk of computation can be done in parallel, and is efficient on modern GPU devices. In addition, we demonstrate that the proposed method can be easily extended by further considering the temporal redundancy, e.g., dynamically skipping less valuable frames. Extensive experiments on five benchmark datasets, i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, demonstrate that our method is significantly more efficient than the competitive baselines. Code is available at https://github.com/blackfeather-wang/AdaFocus.
From computer vision and speech recognition to forecasting trajectories in autonomous vehicles, deep learning approaches are at the forefront of so many domains. Deep learning models are developed using plethora of high-level, generic frameworks and libraries. Running those models on the mobile devices require hardware-aware optimizations and in most cases converting the models to other formats or using a third-party framework. In reality, most of the developed models need to undergo a process of conversion, adaptation, and, in some cases, full retraining to match the requirements and features of the framework that is deploying the model on the target platform. Variety of hardware platforms with heterogeneous computing elements, from wearable devices to high-performance GPU clusters are used to run deep learning models. In this paper, we present the existing challenges, obstacles, and practical solutions towards deploying deep learning models on mobile devices.
The ubiquity of smartphone cameras and IoT cameras, together with the recent boom of deep learning and deep neural networks, proliferate various computer vision driven mobile and IoT applications deployed on the edge. This paper focuses on applications which make soft real time requests to perform inference on their data - they desire prompt responses within designated deadlines, but occasional deadline misses are acceptable. Supporting soft real time applications on a multi-tenant edge server is not easy, since the requests sharing the limited GPU computing resources of an edge server interfere with each other. In order to tackle this problem, we comprehensively evaluate how latency and throughput respond to different GPU execution plans. Based on this analysis, we propose a GPU scheduler, DeepRT, which provides latency guarantee to the requests while maintaining high overall system throughput. The key component of DeepRT, DisBatcher, batches data from different requests as much as possible while it is proven to provide latency guarantee for requests admitted by an Admission Control Module. DeepRT also includes an Adaptation Module which tackles overruns. Our evaluation results show that DeepRT outperforms state-of-the-art works in terms of the number of deadline misses and throughput.
\textit{Graph Neural Network} (GNN) is a promising approach for analyzing graph-structured data that tactfully captures their dependency information via node-level message passing. It has achieved state-of-the-art performances in many tasks, such as node classification, graph matching, clustering, and graph generation. As GNNs operate on non-Euclidean data, their irregular data access patterns cause considerable computational costs and overhead on conventional architectures, such as GPU and CPU. Our analysis shows that GNN adopts a hybrid computing model. The \textit{Aggregation} (or \textit{Message Passing}) phase performs vector additions where vectors are fetched with irregular strides. The \textit{Transformation} (or \textit{Node Embedding}) phase can be either dense or sparse-dense matrix multiplication. In this work, We propose \textit{VersaGNN}, an ultra-efficient, systolic-array-based versatile hardware accelerator that unifies dense and sparse matrix multiplication. By applying this single optimized systolic array to both aggregation and transformation phases, we have significantly reduced chip sizes and energy consumption. We then divide the computing engine into blocked systolic arrays to support the \textit{Strassen}'s algorithm for dense matrix multiplication, dramatically scaling down the number of multiplications and enabling high-throughput computation of GNNs. To balance the workload of sparse-dense matrix multiplication, we also introduced a greedy algorithm to combine sparse sub-matrices of compressed format into condensed ones to reduce computational cycles. Compared with current state-of-the-art GNN software frameworks, \textit{VersaGNN} achieves on average 3712$\times$ speedup with 1301.25$\times$ energy reduction on CPU, and 35.4$\times$ speedup with 17.66$\times$ energy reduction on GPU.
Monocular simultaneous localization and mapping (SLAM) is emerging in advanced driver assistance systems and autonomous driving, because a single camera is cheap and easy to install. Conventional monocular SLAM has two major challenges leading inaccurate localization and mapping. First, it is challenging to estimate scales in localization and mapping. Second, conventional monocular SLAM uses inappropriate mapping factors such as dynamic objects and low-parallax areas in mapping. This paper proposes an improved real-time monocular SLAM that resolves the aforementioned challenges by efficiently using deep learning-based semantic segmentation. To achieve the real-time execution of the proposed method, we apply semantic segmentation only to downsampled keyframes in parallel with mapping processes. In addition, the proposed method corrects scales of camera poses and three-dimensional (3D) points, using estimated ground plane from road-labeled 3D points and the real camera height. The proposed method also removes inappropriate corner features labeled as moving objects and low parallax areas. Experiments with eight video sequences demonstrate that the proposed monocular SLAM system achieves significantly improved and comparable trajectory tracking accuracy, compared to existing state-of-the-art monocular and stereo SLAM systems, respectively. The proposed system can achieve real-time tracking on a standard CPU potentially with a standard GPU support, whereas existing segmentation-aided monocular SLAM does not.
We present a system that allows users to train their own state-of-the-art paraphrastic sentence representations in a variety of languages. We also release trained models for English, Arabic, German, French, Spanish, Russian, Turkish, and Chinese. We train these models on large amounts of data, achieving significantly improved performance from the original papers proposing the methods on a suite of monolingual semantic similarity, cross-lingual semantic similarity, and bitext mining tasks. Moreover, the resulting models surpass all prior work on unsupervised semantic textual similarity, significantly outperforming even BERT-based models like Sentence-BERT (Reimers and Gurevych, 2019). Additionally, our models are orders of magnitude faster than prior work and can be used on CPU with little difference in inference speed (even improved speed over GPU when using more CPU cores), making these models an attractive choice for users without access to GPUs or for use on embedded devices. Finally, we add significantly increased functionality to the code bases for training paraphrastic sentence models, easing their use for both inference and for training them for any desired language with parallel data. We also include code to automatically download and preprocess training data.
Sound sources localization using multichannel signal processing has been a subject of active research for decades. In recent years, the use of deep learning in audio signal processing has allowed to drastically improve performances for machine hearing. This has motivated the scientific community to also develop machine learning strategies for source localization applications. In this paper, we present BeamLearning, a multi-resolution deep learning approach that allows to encode relevant information contained in unprocessed time domain acoustic signals captured by microphone arrays. The use of raw data aims at avoiding simplifying hypothesis that most traditional model-based localization methods rely on. Benefits of its use are shown for realtime sound source 2D-localization tasks in reverberating and noisy environments. Since supervised machine learning approaches require large-sized, physically realistic, precisely labelled datasets, we also developed a fast GPU-based computation of room impulse responses using fractional delays for image source models. A thorough analysis of the network representation and extensive performance tests are carried out using the BeamLearning network with synthetic and experimental datasets. Obtained results demonstrate that the BeamLearning approach significantly outperforms the wideband MUSIC and SRP-PHAT methods in terms of localization accuracy and computational efficiency in presence of heavy measurement noise and reverberation.
Boundary representations (B-reps) using Non-Uniform Rational B-splines (NURBS) are the de facto standard used in CAD, but their utility in deep learning-based approaches is not well researched. We propose a differentiable NURBS module to integrate NURBS representations of CAD models with deep learning methods. We mathematically define the derivatives of the NURBS curves or surfaces with respect to the input parameters (control points, weights, and the knot vector). These derivatives are used to define an approximate Jacobian used for performing the "backward" evaluation to train the deep learning models. We have implemented our NURBS module using GPU-accelerated algorithms and integrated it with PyTorch, a popular deep learning framework. We demonstrate the efficacy of our NURBS module in performing CAD operations such as curve or surface fitting and surface offsetting. Further, we show its utility in deep learning for unsupervised point cloud reconstruction and enforce analysis constraints. These examples show that our module performs better for certain deep learning frameworks and can be directly integrated with any deep-learning framework requiring NURBS.
Object tracking has achieved significant progress over the past few years. However, state-of-the-art trackers become increasingly heavy and expensive, which limits their deployments in resource-constrained applications. In this work, we present LightTrack, which uses neural architecture search (NAS) to design more lightweight and efficient object trackers. Comprehensive experiments show that our LightTrack is effective. It can find trackers that achieve superior performance compared to handcrafted SOTA trackers, such as SiamRPN++ and Ocean, while using much fewer model Flops and parameters. Moreover, when deployed on resource-constrained mobile chipsets, the discovered trackers run much faster. For example, on Snapdragon 845 Adreno GPU, LightTrack runs $12\times$ faster than Ocean, while using $13\times$ fewer parameters and $38\times$ fewer Flops. Such improvements might narrow the gap between academic models and industrial deployments in object tracking task. LightTrack is released at https://github.com/researchmm/LightTrack.
We consider the distributed training of large-scale neural networks that serve as PDE solvers producing full field outputs. We specifically consider neural solvers for the generalized 3D Poisson equation over megavoxel domains. A scalable framework is presented that integrates two distinct advances. First, we accelerate training a large model via a method analogous to the multigrid technique used in numerical linear algebra. Here, the network is trained using a hierarchy of increasing resolution inputs in sequence, analogous to the 'V', 'W', 'F', and 'Half-V' cycles used in multigrid approaches. In conjunction with the multi-grid approach, we implement a distributed deep learning framework which significantly reduces the time to solve. We show the scalability of this approach on both GPU (Azure VMs on Cloud) and CPU clusters (PSC Bridges2). This approach is deployed to train a generalized 3D Poisson solver that scales well to predict output full-field solutions up to the resolution of 512x512x512 for a high dimensional family of inputs.
While learned video codecs have demonstrated great promise, they have yet to achieve sufficient efficiency for practical deployment. In this work, we propose several novel ideas for learned video compression which allow for improved performance for the low-latency mode (I- and P-frames only) along with a considerable increase in computational efficiency. In this setting, for natural videos our approach compares favorably across the entire R-D curve under metrics PSNR, MS-SSIM and VMAF against all mainstream video standards (H.264, H.265, AV1) and all ML codecs. At the same time, our approach runs at least 5x faster and has fewer parameters than all ML codecs which report these figures. Our contributions include a flexible-rate framework allowing a single model to cover a large and dense range of bitrates, at a negligible increase in computation and parameter count; an efficient backbone optimized for ML-based codecs; and a novel in-loop flow prediction scheme which leverages prior information towards more efficient compression. We benchmark our method, which we call ELF-VC (Efficient, Learned and Flexible Video Coding) on popular video test sets UVG and MCL-JCV under metrics PSNR, MS-SSIM and VMAF. For example, on UVG under PSNR, it reduces the BD-rate by 44% against H.264, 26% against H.265, 15% against AV1, and 35% against the current best ML codec. At the same time, on an NVIDIA Titan V GPU our approach encodes/decodes VGA at 49/91 FPS, HD 720 at 19/35 FPS, and HD 1080 at 10/18 FPS.
Scope of Reproducibility Singh et al. [1] point out the dangers of contextual bias in visual recognition datasets. They propose two methods, CAM-based and feature-split, that better recognize an object or attribute in the absence of its typical context while maintaining competitive withincontext accuracy. To verify their performance, we attempted to reproduce all 12 tables in the original paper, including those in the appendix. We also conducted additional experiments to better understand the proposed methods, including increasing the regularization in CAM-based and removing the weighted loss in feature-split. Methodology As the original code was not made available, we implemented the entire pipeline from scratch in PyTorch 1.7.0. Our implementation is based on the paper and email exchanges with the authors. We spent approximately four months reproducing the paper, with the first two months focused on implementing all 10 methods studied in the paper and the next two months focused on reproducing the experiments in the paper and refining our implementation. Total training times for each method ranged from 35–43 hours on COCO-Stuff [2], 22–29 hours on DeepFashion [3], and 7–8 hours on Animals with Attributes [4] on a single RTX 3090 GPU. Results We found that both proposed methods in the original paper help mitigate contextual bias, although for some methods, we could not completely replicate the quantitative results in the paper even after completing an extensive hyperparameter search. For example, on COCO-Stuff, DeepFashion, and UnRel, our feature-split model achieved an increase in accuracy on out-of-context images over the standard baseline, whereas on AwA, we saw a drop in performance. For the proposed CAM-based method, we were able to reproduce the original paperʼs results to within 0.5% mAP. What was easy Overall, it was easy to follow the explanation and reasoning of the experiments. The implementation of most (7 of 10) methods was straightforward, especially after we received additional details from the original authors. What was difficult Since there was no existing code, we spent considerable time and effort re-implementing the entire pipeline from scratch and making sure that most, if not all, training/evaluation details are true to the experiments in the paper. For several methods, we went through many iterations of experiments until we were confident that our implementation was accurate. Communication with original authors We reached out to the authors several times via email to ask for clarifications and additional implementation details. The authors were very responsive to our questions, and we are extremely grateful for their detailed and timely responses.
Object detection is a basic but challenging task in computer vision, which plays a key role in a variety of industrial applications. However, object detectors based on deep learning usually require greater storage requirements and longer inference time, which hinders its practicality seriously. Therefore, a trade-off between effectiveness and efficiency is necessary in practical scenarios. Considering that without constraint of pre-defined anchors, anchor-free detectors can achieve acceptable accuracy and inference speed simultaneously. In this paper, we start from an anchor-free detector called TTFNet, modify the structure of TTFNet and introduce multiple existing tricks to realize effective server and mobile solutions respectively. Since all experiments in this paper are conducted based on PaddlePaddle, we call the model as PAFNet(Paddle Anchor Free Network). For server side, PAFNet can achieve a better balance between effectiveness (42.2% mAP) and efficiency (67.15 FPS) on a single V100 GPU. For moblie side, PAFNet-lite can achieve a better accuracy of (23.9% mAP) and 26.00 ms on Kirin 990 ARM CPU, outperforming the existing state-of-the-art anchor-free detectors by significant margins. Source code is at https://github.com/PaddlePaddle/PaddleDetection.
We release synth1B1, a multi-modal audio corpus consisting of 1 billion 4-second synthesized sounds, paired with the synthesis parameters used to generate them. The dataset is 100x larger than any audio dataset in the literature. We also introduce torchsynth, an open source modular synthesizer that generates the synth1B1 samples on-the-fly at 16200x faster than real-time (714MHz) on a single GPU. Finally, we release two new audio datasets: FM synth timbre and subtractive synth pitch. Using these datasets, we demonstrate new rank-based evaluation criteria for existing audio representations. Finally, we propose a novel approach to synthesizer hyperparameter optimization.
The emergence of CNNs in mainstream deployment has necessitated methods to design and train efficient architectures tailored to maximize the accuracy under diverse hardware & latency constraints. To scale these resource-intensive tasks with an increasing number of deployment targets, Once-For-All (OFA) proposed an approach to jointly train several models at once with a constant training cost. However, this cost remains as high as 40-50 GPU days and also suffers from a combinatorial explosion of sub-optimal model configurations. We seek to reduce this search space -- and hence the training budget -- by constraining search to models close to the accuracy-latency Pareto frontier. We incorporate insights of compound relationships between model dimensions to build CompOFA, a design space smaller by several orders of magnitude. Through experiments on ImageNet, we demonstrate that even with simple heuristics we can achieve a 2x reduction in training time and 216x speedup in model search/extraction time compared to the state of the art, without loss of Pareto optimality! We also show that this smaller design space is dense enough to support equally accurate models for a similar diversity of hardware and latency targets, while also reducing the complexity of the training and subsequent extraction algorithms.
Convolutional neural networks (CNNs) are now the de facto solution for computer vision problems thanks to their impressive results and ease of learning. These networks are composed of layers of connected units called artificial neurons, loosely modeling the neurons in a biological brain. However, their implementation on conventional hardware (CPU/GPU) results in high power consumption, making their integration on embedded systems difficult. In a car for example, embedded algorithms have very high constraints in term of energy, latency and accuracy. To design more efficient computer vision algorithms, we propose to follow an end-to-end biologically inspired approach using event cameras and spiking neural networks (SNNs). Event cameras output asynchronous and sparse events, providing an incredibly efficient data source, but processing these events with synchronous and dense algorithms such as CNNs does not yield any significant benefits. To address this limitation, we use spiking neural networks (SNNs), which are more biologically realistic neural networks where units communicate using discrete spikes. Due to the nature of their operations, they are hardware friendly and energy-efficient, but training them still remains a challenge. Our method enables the training of sparse spiking convolutional neural networks directly on event data, using the popular deep learning framework PyTorch. The performances in terms of accuracy, sparsity and training time on the popular DVS128 Gesture Dataset make it possible to use this bio-inspired approach for the future embedding of real-time applications on low-power neuromorphic hardware.
Recently, large-scale transformer-based models have been proven to be effective over various tasks across many domains. Nevertheless, applying them in industrial production requires tedious and heavy works to reduce inference costs. To fill such a gap, we introduce a scalable inference solution: Easy and Efficient Transformer (EET), including a series of transformer inference optimization at the algorithm and implementation levels. First, we design highly optimized kernels for long inputs and large hidden sizes. Second, we propose a flexible CUDA memory manager to reduce the memory footprint when deploying a large model. Compared with the state-of-the-art transformer inference library (Faster Transformer v4.0), EET can achieve an average of 1.40-4.20x speedup on the transformer decoder layer with an A100 GPU
We introduce CryptGPU, a system for privacy-preserving machine learning that implements all operations on the GPU (graphics processing unit). Just as GPUs played a pivotal role in the success of modern deep learning, they are also essential for realizing scalable privacy-preserving deep learning. In this work, we start by introducing a new interface to losslessly embed cryptographic operations over secret-shared values (in a discrete domain) into floating-point operations that can be processed by highly-optimized CUDA kernels for linear algebra. We then identify a sequence of "GPU-friendly" cryptographic protocols to enable privacy-preserving evaluation of both linear and non-linear operations on the GPU. Our microbenchmarks indicate that our private GPU-based convolution protocol is over 150x faster than the analogous CPU-based protocol; for non-linear operations like the ReLU activation function, our GPU-based protocol is around 10x faster than its CPU analog. With CryptGPU, we support private inference and private training on convolutional neural networks with over 60 million parameters as well as handle large datasets like ImageNet. Compared to the previous state-of-the-art, when considering large models and datasets, our protocols achieve a 2x to 8x improvement in private inference and a 6x to 36x improvement for private training. Our work not only showcases the viability of performing secure multiparty computation (MPC) entirely on the GPU to enable fast privacy-preserving machine learning, but also highlights the importance of designing new MPC primitives that can take full advantage of the GPU's computing capabilities.
Graph neural networks (GNNs), an emerging deep learning model class, can extract meaningful representations from highly expressive graph-structured data and are therefore gaining popularity for wider ranges of applications. However, current GNNs suffer from the poor performance of their sparse-dense matrix multiplication (SpMM) operator, even when using powerful GPUs. Our analysis shows that 95% of the inference time could be spent on SpMM when running popular GNN models on NVIDIA's advanced V100 GPU. Such SpMM performance bottleneck hinders GNNs' applicability to large-scale problems or the development of more sophisticated GNN models. To address this inference time bottleneck, we introduce ES-SpMM, a cache-first edge sampling mechanism and codesigned SpMM kernel. ES-SpMM uses edge sampling to downsize the graph to fit into GPU's shared memory. It thus reduces the computation cost and improves SpMM's cache locality. To evaluate ES-SpMM's performance, we integrated it with a popular GNN framework, DGL, and tested it using representative GNN models and datasets. Our results show that ES-SpMM outperforms the highly optimized cuSPARSE SpMM kernel by up to 4.35x with no accuracy loss and by 45.3x with less than a 1% accuracy loss.