This paper presents algorithms for parallelization of inference in hidden Markov models (HMMs). In particular, we propose parallel backward-forward type of filtering and smoothing algorithm as well as parallel Viterbi-type maximum-a-posteriori (MAP) algorithm. We define associative elements and operators to pose these inference problems as parallel-prefix-sum computations in sum-product and max-product algorithms and parallelize them using parallel-scan algorithms. The advantage of the proposed algorithms is that they are computationally efficient in HMM inference problems with long time horizons. We empirically compare the performance of the proposed methods to classical methods on a highly parallel graphical processing unit (GPU).
Instance segmentation has gained recently huge attention in various computer vision applications. It aims at providing different IDs to different object of the scene, even if they belong to the same class. This is useful in various scenarios, especially in occlusions. Instance segmentation is usually performed as a two-stage pipeline. First, an object is detected, then semantic segmentation within the detected box area. This process involves costly up-sampling, especially for the segmentation part. Moreover, for some applications, such as LiDAR point clouds and aerial object detection, it is often required to predict oriented boxes, which add extra complexity to the two-stage pipeline. In this paper, we propose Insta-YOLO, a novel one-stage end-to-end deep learning model for real-time instance segmentation. The proposed model is inspired by the YOLO one-shot object detector, with the box regression loss is replaced with polynomial regression in the localization head. This modification enables us to skip the segmentation up-sampling decoder altogether and produces the instance segmentation contour from the polynomial output coefficients. In addition, this architecture is a natural fit for oriented objects. We evaluate our model on three datasets, namely, Carnva, Cityscapes and Airbus. The results show our model achieves competitive accuracy in terms of mAP with significant improvement in speed by 2x on GTX-1080 GPU.
In the last decades, extreme classification has become an essential topic for deep learning. It has achieved great success in many areas, especially in computer vision and natural language processing (NLP). However, it is very challenging to train a deep model with millions of classes due to the memory and computation explosion in the last output layer. In this paper, we propose a large-scale training system to address these challenges. First, we build a hybrid parallel training framework to make the training process feasible. Second, we propose a novel softmax variation named KNN softmax, which reduces both the GPU memory consumption and computation costs and improves the throughput of training. Then, to eliminate the communication overhead, we propose a new overlapping pipeline and a gradient sparsification method. Furthermore, we design a fast continuous convergence strategy to reduce total training iterations by adaptively adjusting learning rate and updating model parameters. With the help of all the proposed methods, we gain 3.9$\times$ throughput of our training system and reduce almost 60\% of training iterations. The experimental results show that using an in-house 256 GPUs cluster, we could train a classifier of 100 million classes on Alibaba Retail Product Dataset in about five days while achieving a comparable accuracy with the naive softmax training process.
We present an ensemble prediction system using a Deep Learning Weather Prediction (DLWP) model that recursively predicts key atmospheric variables with six-hour time resolution. This model uses convolutional neural networks (CNNs) on a cubed sphere grid to produce global forecasts. The approach is computationally efficient, requiring just three minutes on a single GPU to produce a 320-member set of six-week forecasts at 1.4{\deg} resolution. Ensemble spread is primarily produced by randomizing the CNN training process to create a set of 32 DLWP models with slightly different learned weights. Although our DLWP model does not forecast precipitation, it does forecast total column water vapor, and it gives a reasonable 4.5-day deterministic forecast of Hurricane Irma. In addition to simulating mid-latitude weather systems, it spontaneously generates tropical cyclones in a one-year free-running simulation. Averaged globally and over a two-year test set, the ensemble mean RMSE retains skill relative to climatology beyond two-weeks, with anomaly correlation coefficients remaining above 0.6 through six days. Our primary application is to subseasonal-to-seasonal (S2S) forecasting at lead times from two to six weeks. Current forecast systems have low skill in predicting one- or 2-week-average weather patterns at S2S time scales. The continuous ranked probability score (CRPS) and the ranked probability skill score (RPSS) show that the DLWP ensemble is only modestly inferior in performance to the European Centre for Medium Range Weather Forecasts (ECMWF) S2S ensemble over land at lead times of 4 and 5-6 weeks. At shorter lead times, the ECMWF ensemble performs better than DLWP.
We have developed several autotuning benchmarks in CUDA that take into account performance-relevant source-code parameters and reach near peak-performance on various GPU architectures. We have used them during the development and evaluation of a novel search method for tuning space proposed in [1]. With our framework Kernel Tuning Toolkit, freely available at Github, we measured computation times and hardware performance counters on several GPUs for the complete tuning spaces of five benchmarks. These data, which we provide here, might benefit research of search algorithms for the tuning spaces of GPU codes or research of relation between applied code optimization, hardware performance counters, and GPU kernels' performance. Moreover, we describe the scripts we used for robust evaluation of our searcher and comparison to others in detail. In particular, the script that simulates the tuning, i.e., replaces time-demanding compiling and executing the tuned kernels with a quick reading of the computation time from our measured data, makes it possible to inspect the convergence of tuning search over a large number of experiments. These scripts, freely available with our other codes, make it easier to experiment with search algorithms and compare them in a robust way. During our research, we generated models for predicting values of performance counters from values of tuning parameters of our benchmarks. Here, we provide the models themselves and describe the scripts we implemented for their training. These data might benefit researchers who want to reproduce or build on our research.
Nowadays, GPU accelerators are commonly used to speed up general-purpose computing tasks on a variety of hardware. However, due to the diversity of GPU architectures and processed data, optimization of codes for a particular type of hardware and specific data characteristics can be extremely challenging. The autotuning of performance-relevant source-code parameters allows for automatic optimization of applications and keeps their performance portable. Although the autotuning process typically results in code speed-up, searching the tuning space can bring unacceptable overhead if (i) the tuning space is vast and full of poorly-performing implementations, or (ii) the autotuning process has to be repeated frequently because of changes in processed data or migration to different hardware. In this paper, we introduce a novel method for searching tuning spaces. The method takes advantage of collecting hardware performance counters (also known as profiling counters) during empirical tuning. Those counters are used to navigate the searching process towards faster implementations. The method requires the tuning space to be sampled on any GPU. It builds a problem-specific model, which can be used during autotuning on various, even previously unseen inputs or GPUs. Using a set of five benchmarks, we experimentally demonstrate that our method can speed up autotuning when an application needs to be ported to different hardware or when it needs to process data with different characteristics. We also compared our method to state of the art and show that our method is superior in terms of the number of searching steps and typically outperforms other searches in terms of convergence time.
We present a SNN simulator which scales to millions of neurons, billions of synapses, and 8 GPUs. This is made possible by 1) a novel, cache-aware spike transmission algorithm 2) a model parallel multi-GPU distribution scheme and 3) a static, yet very effective load balancing strategy. The simulator further features an easy to use API and the ability to create custom models. We compare the proposed simulator against two state of the art ones on a series of benchmarks using three well-established models. We find that our simulator is faster, consumes less memory, and scales linearly with the number of GPUs.
The ability to distinguish between the self and the background is of paramount importance for robotic tasks. The particular case of hands, as the end effectors of a robotic system that more often enter into contact with other elements of the environment, must be perceived and tracked with precision to execute the intended tasks with dexterity and without colliding with obstacles. They are fundamental for several applications, from Human-Robot Interaction tasks to object manipulation. Modern humanoid robots are characterized by high number of degrees of freedom which makes their forward kinematics models very sensitive to uncertainty. Thus, resorting to vision sensing can be the only solution to endow these robots with a good perception of the self, being able to localize their body parts with precision. In this paper, we propose the use of a Convolution Neural Network (CNN) to segment the robot hand from an image in an egocentric view. It is known that CNNs require a huge amount of data to be trained. To overcome the challenge of labeling real-world images, we propose the use of simulated datasets exploiting domain randomization techniques. We fine-tuned the Mask-RCNN network for the specific task of segmenting the hand of the humanoid robot Vizzy. We focus our attention on developing a methodology that requires low amounts of data to achieve reasonable performance while giving detailed insight on how to properly generate variability in the training dataset. Moreover, we analyze the fine-tuning process within the complex model of Mask-RCNN, understanding which weights should be transferred to the new task of segmenting robot hands. Our final model was trained solely on synthetic images and achieves an average IoU of 82% on synthetic validation data and 56.3% on real test data. These results were achieved with only 1000 training images and 3 hours of training time using a single GPU.
Deep reinforcement learning (RL) has made groundbreaking advancements in robotics, data center management and other applications. Unfortunately, system-level bottlenecks in RL workloads are poorly understood; we observe fundamental structural differences in RL workloads that make them inherently less GPU-bound than supervised learning (SL). To explain where training time is spent in RL workloads, we propose RL-Scope, a cross-stack profiler that scopes low-level CPU/GPU resource usage to high-level algorithmic operations, and provides accurate insights by correcting for profiling overhead. Using RL-Scope, we survey RL workloads across its major dimensions including ML backend, RL algorithm, and simulator. For ML backends, we explain a $2.3\times$ difference in runtime between equivalent PyTorch and TensorFlow algorithm implementations, and identify a bottleneck rooted in overly abstracted algorithm implementations. For RL algorithms and simulators, we show that on-policy algorithms are at least $3.5\times$ more simulation-bound than off-policy algorithms. Finally, we profile a scale-up workload and demonstrate that GPU utilization metrics reported by commonly used tools dramatically inflate GPU usage, whereas RL-Scope reports true GPU-bound time. RL-Scope is an open-source tool available at https://github.com/UofT-EcoSystem/rlscope .
Adversarial patches pose a realistic threat model for physical world attacks on autonomous systems via their perception component. Autonomous systems in safety-critical domains such as automated driving should thus contain a fail-safe fallback component that combines certifiable robustness against patches with efficient inference while maintaining high performance on clean inputs. We propose BagCert, a novel combination of model architecture and certification procedure that allows efficient certification. We derive a loss that enables end-to-end optimization of certified robustness against patches of different sizes and locations. On CIFAR10, BagCert certifies 10.000 examples in 43 seconds on a single GPU and obtains 86% clean and 60% certified accuracy against 5x5 patches.
Many compelling video processing effects can be achieved if per-pixel depth information and 3D camera calibrations are known. However, the success of such methods is highly dependent on the accuracy of this "scene-space" information. We present a novel, sampling-based framework for processing video that enables high-quality scene-space video effects in the presence of inevitable errors in depth and camera pose estimation. Instead of trying to improve the explicit 3D scene representation, the key idea of our method is to exploit the high redundancy of approximate scene information that arises due to most scene points being visible multiple times across many frames of video. Based on this observation, we propose a novel pixel gathering and filtering approach. The gathering step is general and collects pixel samples in scene-space, while the filtering step is application-specific and computes a desired output video from the gathered sample sets. Our approach is easily parallelizable and has been implemented on GPU, allowing us to take full advantage of large volumes of video data and facilitating practical runtimes on HD video using a standard desktop computer. Our generic scene-space formulation is able to comprehensively describe a multitude of video processing applications such as denoising, deblurring, super resolution, object removal, computational shutter functions, and other scene-space camera effects. We present results for various casually captured, hand-held, moving, compressed, monocular videos depicting challenging scenes recorded in uncontrolled environments.
Driven by the tremendous effort in researching novel deep learning (DL) algorithms, the training cost of developing new models increases staggeringly in recent years. We analyze GPU cluster usage statistics from a top research institute for more insights into the hardware efficiency achieved by typical DL training jobs. Our study reveals that single-accelerator training jobs can dominate the cluster-wide resource consumption when launched repetitively (e.g., for hyper-parameter tuning) while severely under-utilizing the hardware. Fortunately, we observe that such workloads have the following unique characteristics: (i) the models among jobs often have the same types of operators with the same shapes, and (ii) the inter-model horizontal fusion of such operators is mathematically equivalent to other already well-optimized operators. Thus, to help DL researchers and practitioners effectively improve the hardware utilization of their novel DL training workloads, we propose Horizontally Fused Training Array (HFTA). HFTA is a new DL framework extension library that horizontally fuses the models from different repetitive jobs deeply down to operators and then trains them simultaneously on a shared accelerator. To show the generality of our solution, we apply HFTA to six DL models training on state-of-the-art accelerators (GPUs and TPUs). Our results indicate that HFTA is highly effective in improving hardware utilization and achieves up to $15.1 \times$ higher training throughput vs. the standard practice of running each job on a separate accelerator.
On-device machine learning is becoming a reality thanks to the availability of powerful hardware and model compression techniques. Typically, these models are pretrained on large GPU clusters and have enough parameters to generalise across a wide variety of inputs. In this work, we observe that a much smaller, personalised model can be employed to fit a specific scenario, resulting in both higher accuracy and faster execution. Nevertheless, on-device training is extremely challenging, imposing excessive computational and memory requirements even for flagship smartphones. At the same time, on-device data availability might be limited and samples are most frequently unlabelled. To this end, we introduce PersEPhonEE, a framework that attaches early exits on the model and personalises them on-device. These allow the model to progressively bypass a larger part of the computation as more personalised data become available. Moreover, we introduce an efficient on-device algorithm that trains the early exits in a semi-supervised manner at a fraction of the whole network's personalisation time. Results show that PersEPhonEE boosts accuracy by up to 15.9% while dropping the training cost by up to 2.2x and inference latency by 2.2-3.2x on average for the same accuracy, depending on the availability of labels on-device.
With the rapid adoption of machine learning (ML), a number of domains now use the approach of fine tuning models which were pre-trained on a large corpus of data. However, our experiments show that even fine-tuning on models like BERT can take many hours even when using modern accelerators like GPUs. While prior work proposes limiting the number of layers that are fine-tuned, e.g., freezing all layers but the last layer, we find that such static approaches lead to reduced accuracy. We propose, AutoFreeze, a system that uses an adaptive approach to choose which layers are trained and show how this can accelerate model fine-tuning while preserving accuracy. We also develop mechanisms to enable efficient caching of intermediate activations which can reduce the forward computation time when performing fine-tuning. We extend AutoFreeze to perform distributed fine-tuning and design two execution modes that minimize cost and running time respectively. Our evaluation on ten NLP tasks shows that AutoFreeze, with caching enabled, can improve fine-tuning on a single GPU by up to 2.55x. On a 64 GPU cluster, for fine-tuning on the AG's news dataset, AutoFreeze is able to achieve up to 4.38x speedup when optimizing for end-to-end training time and 5.03x reduction in total cost when optimizing for efficiency, without affecting model accuracy.
Wide dynamic range (WDR) image tone mapping is in high demand in many applications like film production, security monitoring, and photography. It is especially crucial for mobile devices because most of the images taken today are from mobile phones, hence such technology is highly demanded in the consumer market of mobile devices and is essential for a good customer experience. However, high-quality and high-performance WDR image tone mapping implementations are rarely found in the mobile-end. In this paper, we introduce a high performance, mobile-end WDR image tone mapping implementation. It leverages the tone mapping results of multiple receptive fields and calculates a suitable value for each pixel. The utilization of integral image and integral histogram significantly reduce the required computation. Moreover, GPU parallel computation is used to increase the processing speed. The experimental results indicate that our implementation can process a high-resolution WDR image within a second on mobile devices and produce appealing image quality.
Quantitative MR imaging is increasingly favoured for its richer information content and standardised measures. However, computing quantitative parameter maps, such as those encoding longitudinal relaxation rate (R1), apparent transverse relaxation rate (R2*) or magnetisation-transfer saturation (MTsat), involves inverting a highly non-linear function. Many methods for deriving parameter maps assume perfect measurements and do not consider how noise is propagated through the estimation procedure, resulting in needlessly noisy maps. Instead, we propose a probabilistic generative (forward) model of the entire dataset, which is formulated and inverted to jointly recover (log) parameter maps with a well-defined probabilistic interpretation (e.g., maximum likelihood or maximum a posteriori). The second order optimisation we propose for model fitting achieves rapid and stable convergence thanks to a novel approximate Hessian. We demonstrate the utility of our flexible framework in the context of recovering more accurate maps from data acquired using the popular multi-parameter mapping protocol. We also show how to incorporate a joint total variation prior to further decrease the noise in the maps, noting that the probabilistic formulation allows the uncertainty on the recovered parameter maps to be estimated. Our implementation uses a PyTorch backend and benefits from GPU acceleration. It is available at https://github.com/balbasty/nitorch.
Cycling is a promising sustainable mode for commuting and leisure in cities, however, the fear of getting hit or fall reduces its wide expansion as a commuting mode. In this paper, we introduce a novel method called CyclingNet for detecting cycling near misses from video streams generated by a mounted frontal camera on a bike regardless of the camera position, the conditions of the built, the visual conditions and without any restrictions on the riding behaviour. CyclingNet is a deep computer vision model based on convolutional structure embedded with self-attention bidirectional long-short term memory (LSTM) blocks that aim to understand near misses from both sequential images of scenes and their optical flows. The model is trained on scenes of both safe rides and near misses. After 42 hours of training on a single GPU, the model shows high accuracy on the training, testing and validation sets. The model is intended to be used for generating information that can draw significant conclusions regarding cycling behaviour in cities and elsewhere, which could help planners and policy-makers to better understand the requirement of safety measures when designing infrastructure or drawing policies. As for future work, the model can be pipelined with other state-of-the-art classifiers and object detectors simultaneously to understand the causality of near misses based on factors related to interactions of road-users, the built and the natural environments.
Deep learning researchers and practitioners usually leverage GPUs to help train their deep neural networks (DNNs) faster. However, choosing which GPU to use is challenging both because (i) there are many options, and (ii) users grapple with competing concerns: maximizing compute performance while minimizing costs. In this work, we present a new practical technique to help users make informed and cost-efficient GPU selections: make performance predictions with the help of a GPU that the user already has. Our technique exploits the observation that, because DNN training consists of repetitive compute steps, predicting the execution time of a single iteration is usually enough to characterize the performance of an entire training process. We make predictions by scaling the execution time of each operation in a training iteration from one GPU to another using either (i) wave scaling, a technique based on a GPU's execution model, or (ii) pre-trained multilayer perceptrons. We implement our technique into a Python library called Habitat and find that it makes accurate iteration execution time predictions (with an average error of 11.8%) on ResNet-50, Inception v3, the Transformer, GNMT, and DCGAN across six different GPU architectures. Habitat supports PyTorch, is easy to use, and is open source.
Accuracy predictor is a key component in Neural Architecture Search (NAS) for ranking architectures. Building a high-quality accuracy predictor usually costs enormous computation. To address this issue, instead of using an accuracy predictor, we propose a novel zero-shot index dubbed Zen-Score to rank the architectures. The Zen-Score represents the network expressivity and positively correlates with the model accuracy. The calculation of Zen-Score only takes a few forward inferences through a randomly initialized network, without training network parameters. Built upon the Zen-Score, we further propose a new NAS algorithm, termed as Zen-NAS, by maximizing the Zen-Score of the target network under given inference budgets. Within less than half GPU day, Zen-NAS is able to directly search high performance architectures in a data-free style. Comparing with previous NAS methods, the proposed Zen-NAS is magnitude times faster on multiple server-side and mobile-side GPU platforms with state-of-the-art accuracy on ImageNet. Our source code and pre-trained models are released on https://github.com/idstcv/ZenNAS.
This work introduces a new preprocessing step for object detection applicable to UAV bird's eye view imagery, which we call Adaptive Resizing. By design, it helps alleviate the challenges coming with the vast variances in objects' scales, naturally inherent to UAV data sets. Furthermore, it improves inference speed by two to three times on average. We test this extensively on UAVDT, VisDrone, and on a new data set we captured ourselves and achieve consistent improvements while being considerably faster. Moreover, we show how to apply this method to generic UAV object detection tasks. Additionally, we successfully test our approach on a height transfer task where we train on some interval of altitudes and test on a different one. Furthermore, we introduce a small, fast detector meant for deployment to an embedded GPU. Code will be made publicly available on our website.