Training deep convolutional neural networks such as VGG and ResNet by gradient descent is an expensive exercise requiring specialized hardware such as GPUs. Recent works have examined the possibility of approximating the gradient computation while maintaining the same convergence properties. While promising, the approximations only work on relatively small datasets such as MNIST. They also fail to achieve real wall-clock speedups due to lack of efficient GPU implementations of the proposed approximation methods. In this work, we explore three alternative methods to approximate gradients, with an efficient GPU kernel implementation for one of them. We achieve wall-clock speedup with ResNet-20 and VGG-19 on the CIFAR-10 dataset upwards of 7%, with a minimal loss in validation accuracy.
Supervised learning is the workhorse for regression and classification tasks, but the standard approach presumes ground truth for every measurement. In real world applications, limitations due to expense or general in-feasibility due to the specific application are common. In the context of agriculture applications, yield monitoring is one such example where simple-physics based measurements such as volume or force-impact have been used to quantify mass flow, which incur error due to sensor calibration. By utilizing semi-supervised deep learning with gradient aggregation and a sequence of images, in this work we can accurately estimate a physical quantity (mass) with complex data structures and sparse ground truth. Using a vision system capturing images of a sugarcane elevator and running bamboo under controlled testing as a surrogate material to harvesting sugarcane, mass is accurately predicted from images by training a DNN using only final load weights. The DNN succeeds in capturing the complex density physics of random stacking of slender rods internally as part of the mass prediction model, and surpasses older volumetric-based methods for mass prediction. Furthermore, by incorporating knowledge about the system physics through the DNN architecture and penalty terms, improvements in prediction accuracy and stability, as well as faster learning are obtained. It is shown that the classic nonlinear regression optimization can be reformulated with an aggregation term with some independence assumptions to achieve this feat. Since the number of images for any given run are too large to fit on typical GPU vRAM, an implementation is shown that compensates for the limited memory but still achieve fast training times. The same approach presented herein could be applied to other applications like yield monitoring on grain combines or other harvesters using vision or other instrumentation.
The shear volumes of data generated from earth observation and remote sensing technologies continue to make major impact; leaping key geospatial applications into the dual data and compute intensive era. As a consequence, this rapid advancement poses new computational and data processing challenges. We implement a novel remote sensing data flow (RESFlow) for advanced machine learning and computing with massive amounts of remotely sensed imagery. The core contribution is partitioning massive amount of data based on the spectral and semantic characteristics for distributed imagery analysis. RESFlow takes advantage of both a unified analytics engine for large-scale data processing and the availability of modern computing hardware to harness the acceleration of deep learning inference on expansive remote sensing imagery. The framework incorporates a strategy to optimize resource utilization across multiple executors assigned to a single worker. We showcase its deployment across computationally and data-intensive on pixel-level labeling workloads. The pipeline invokes deep learning inference at three stages; during deep feature extraction, deep metric mapping, and deep semantic segmentation. The tasks impose compute intensive and GPU resource sharing challenges motivating for a parallelized pipeline for all execution steps. By taking advantage of Apache Spark, Nvidia DGX1, and DGX2 computing platforms, we demonstrate unprecedented compute speed-ups for deep learning inference on pixel labeling workloads; processing 21,028~Terrabytes of imagery data and delivering an output maps at area rate of 5.245sq.km/sec, amounting to 453,168 sq.km/day - reducing a 28 day workload to 21~hours.
As designing appropriate Convolutional Neural Network (CNN) architecture in the context of a given application usually involves heavy human works or numerous GPU hours, the research community is soliciting the architecture-neutral CNN structures, which can be easily plugged into multiple mature architectures to improve the performance on our real-world applications. We propose Asymmetric Convolution Block (ACB), an architecture-neutral structure as a CNN building block, which uses 1D asymmetric convolutions to strengthen the square convolution kernels. For an off-the-shelf architecture, we replace the standard square-kernel convolutional layers with ACBs to construct an Asymmetric Convolutional Network (ACNet), which can be trained to reach a higher level of accuracy. After training, we equivalently convert the ACNet into the same original architecture, thus requiring no extra computations anymore. We have observed that ACNet can improve the performance of various models on CIFAR and ImageNet by a clear margin. Through further experiments, we attribute the effectiveness of ACB to its capability of enhancing the model's robustness to rotational distortions and strengthening the central skeleton parts of square convolution kernels.
Recent studies from several hyperscalars pinpoint to embedding layers as the most memory-intensive deep learning (DL) algorithm being deployed in today's datacenters. This paper addresses the memory capacity and bandwidth challenges of embedding layers and the associated tensor operations. We present our vertically integrated hardware/software co-design, which includes a custom DIMM module enhanced with near-data processing cores tailored for DL tensor operations. These custom DIMMs are populated inside a GPU-centric system interconnect as a remote memory pool, allowing GPUs to utilize for scalable memory bandwidth and capacity expansion. A prototype implementation of our proposal on real DL systems shows an average 6.2-17.6x performance improvement on state-of-the-art recommender systems.
We present a learning model that makes full use of boundary information for salient object segmentation. Specifically, we come up with a novel loss function, i.e., Contour Loss, which leverages object contours to guide models to perceive salient object boundaries. Such a boundary-aware network can learn boundary-wise distinctions between salient objects and background, hence effectively facilitating the saliency detection. Yet the Contour Loss emphasizes on the local saliency. We further propose the hierarchical global attention module (HGAM), which forces the model hierarchically attend to global contexts, thus captures the global visual saliency. Comprehensive experiments on six benchmark datasets show that our method achieves superior performance over state-of-the-art ones. Moreover, our model has a real-time speed of 26 fps on a TITAN X GPU.
The evolution of MobileNets has laid a solid foundation for neural network applications on mobile end. With the latest MobileNetV3, neural architecture search again claimed its supremacy in network design. Unfortunately, till today all mobile methods mainly focus on CPU latencies instead of GPU, the latter, however, is much preferred in practice for it has faster speed, lower overhead and less interference. Bearing the target hardware in mind, we propose the first Mobile GPU-Aware (MoGA) neural architecture search in order to be precisely tailored for real-world applications. Further, the ultimate objective to devise a mobile network lies in achieving better performance by maximizing the utilization of bounded resources. Urging higher capability while restraining time consumption is not reconcilable. We alleviate the tension by weighted evolution techniques. Moreover, we encourage increasing the number of parameters for higher representational power. With 200x fewer GPU days than MnasNet, we obtain a series of models that outperform MobileNetV3 under the similar latency constraints, i.e., MoGA-A achieves 75.9% top-1 accuracy on ImageNet, MoGA-B meets 75.5% which costs only 0.5 ms more on mobile GPU. MoGA-C best attests GPU-awareness by reaching 75.3% and being slower on CPU but faster on GPU.The models and test code is made available here https://github.com/xiaomi-automl/MoGA.
Dense depth perception is critical for autonomous driving and other robotics applications. However, modern LiDAR sensors only provide sparse depth measurement. It is thus necessary to complete the sparse LiDAR data, where a synchronized guidance RGB image is often used to facilitate this completion. Many neural networks have been designed for this task. However, they often na\"{\i}vely fuse the LiDAR data and RGB image information by performing feature concatenation or element-wise addition. Inspired by the guided image filtering, we design a novel guided network to predict kernel weights from the guidance image. These predicted kernels are then applied to extract the depth image features. In this way, our network generates content-dependent and spatially-variant kernels for multi-modal feature fusion. Dynamically generated spatially-variant kernels could lead to prohibitive GPU memory consumption and computation overhead. We further design a convolution factorization to reduce computation and memory consumption. The GPU memory reduction makes it possible for feature fusion to work in multi-stage scheme. We conduct comprehensive experiments to verify our method on real-world outdoor, indoor and synthetic datasets. Our method produces strong results. It outperforms state-of-the-art methods on the NYUv2 dataset and ranks 1st on the KITTI depth completion benchmark at the time of submission. It also presents strong generalization capability under different 3D point densities, various lighting and weather conditions as well as cross-dataset evaluations. The code will be released for reproduction.
Visual Place Recognition (VPR) is a fundamental yet challenging task for small Unmanned Aerial Vehicle (UAV). The core reasons are the extreme viewpoint changes, and limited computational power onboard a UAV which restricts the applicability of robust but computation intensive state-of-the-art VPR methods. In this context, a viable approach is to use local image descriptors for performing VPR as these can be computed relatively efficiently without the need of any special hardware, such as a GPU. However, the choice of a local feature descriptor is not trivial and calls for a detailed investigation as there is a trade-off between VPR accuracy and the required computational effort. To fill this research gap, this paper examines the performance of several state-of-the-art local feature descriptors, both from accuracy and computational perspectives, specifically for VPR application utilizing standard aerial datasets. The presented results confirm that a trade-off between accuracy and computational effort is inevitable while executing VPR on resource-constrained hardware.
Deploying deep learning (DL) models across multiple compute devices to train large and complex models continues to grow in importance because of the demand for faster and more frequent training. Data parallelism (DP) is the most widely used parallelization strategy, but as the number of devices in data parallel training grows, so does the communication overhead between devices. Additionally, a larger aggregate batch size per step leads to statistical efficiency loss, i.e., a larger number of epochs are required to converge to a desired accuracy. These factors affect overall training time and beyond a certain number of devices, the speedup from leveraging DP begins to scale poorly. In addition to DP, each training step can be accelerated by exploiting model parallelism (MP). This work explores hybrid parallelization, where each data parallel worker is comprised of more than one device, across which the model dataflow graph (DFG) is split using MP. We show that at scale, hybrid training will be more effective at minimizing end-to-end training time than exploiting DP alone. We project that for Inception-V3, GNMT, and BigLSTM, the hybrid strategy provides an end-to-end training speedup of at least 26.5%, 8%, and 22% respectively compared to what DP alone can achieve at scale.
Artificial Neural networks (ANNs) are powerful computing systems employed for various applications due to their versatility to generalize and to respond to unexpected inputs/patterns. However, implementations of ANNs for safety-critical systems might lead to failures, which are hardly predicted in the design phase since ANNs are highly parallel and their parameters are hardly interpretable. Here we develop and evaluate a novel symbolic software verification framework based on incremental bounded model checking (BMC) to check for adversarial cases and coverage methods in multi-layer perceptron (MLP). In particular, we further develop the efficient SMT-based Context-Bounded Model Checker for Graphical Processing Units (ESBMC-GPU) in order to ensure the reliability of certain safety properties in which safety-critical systems can fail and make incorrect decisions, thereby leading to unwanted material damage or even put lives in danger. This paper marks the first symbolic verification framework to reason over ANNs implemented in CUDA. Our experimental results show that our approach implemented in ESBMC-GPU can successfully verify safety properties and covering methods in ANNs and correctly generate 28 adversarial cases in MLPs.
In recent years, deep-learning-based visual object trackers have been studied thoroughly, but handling occlusions and/or rapid motion of the target remains challenging. In this work, we argue that conditioning on the natural language (NL) description of a target provides information for longer-term invariance, and thus helps cope with typical tracking challenges. However, deriving a formulation to combine the strengths of appearance-based tracking with the language modality is not straightforward. We propose a novel deep tracking-by-detection formulation that can take advantage of NL descriptions. Regions that are related to the given NL description are generated by a proposal network during the detection phase of the tracker. Our LSTM based tracker then predicts the update of the target from regions proposed by the NL based detection phase. In benchmarks, our method is competitive with state of the art trackers, while it outperforms all other trackers on targets with unambiguous and precise language annotations. It also beats the state-of-the-art NL tracker when initializing without a bounding box. Our method runs at over 30 fps on a single GPU.
Training deep learning models is compute-intensive and there is an industry-wide trend towards hardware specialization to improve performance. To systematically benchmark deep learning platforms, we introduce ParaDnn, a parameterized benchmark suite for deep learning that generates end-to-end models for fully connected (FC), convolutional (CNN), and recurrent (RNN) neural networks. Along with six real-world models, we benchmark Google's Cloud TPU v2/v3, NVIDIA's V100 GPU, and an Intel Skylake CPU platform. We take a deep dive into TPU architecture, reveal its bottlenecks, and highlight valuable lessons learned for future specialized system design. We also provide a thorough comparison of the platforms and find that each has unique strengths for some types of models. Finally, we quantify the rapid performance improvements that specialized software stacks provide for the TPU and GPU platforms.
Deep learning techniques, namely convolutional neural networks (CNN), have previously been adapted to select gamma-ray events in the TAIGA experiment, having achieved a good quality of selection as compared with the conventional Hillas approach. Another important task for the TAIGA data analysis was also solved with CNN: gamma-ray energy estimation showed some improvement in comparison with the conventional method based on the Hillas analysis. Furthermore, our software was completely redeveloped for the graphics processing unit (GPU), which led to significantly faster calculations in both of these tasks. All the results have been obtained with the simulated data of TAIGA Monte Carlo software; their experimental confirmation is envisaged for the near future.
We address a learning-to-normalize problem by proposing Switchable Normalization (SN), which learns to select different normalizers for different normalization layers of a deep neural network. SN employs three distinct scopes to compute statistics (means and variances) including a channel, a layer, and a minibatch. SN switches between them by learning their importance weights in an end-to-end manner. It has several good properties. First, it adapts to various network architectures and tasks. Second, it is robust to a wide range of batch sizes, maintaining high performance even when small minibatch is presented (e.g. 2 images/GPU). Third, SN does not have sensitive hyper-parameter, unlike group normalization that searches the number of groups as a hyper-parameter. Without bells and whistles, SN outperforms its counterparts on various challenging benchmarks, such as ImageNet, COCO, CityScapes, ADE20K, MegaFace, and Kinetics. Analyses of SN are also presented to answer the following three questions: (a) Is it useful to allow each normalization layer to select its own normalizer? (b) What impacts the choices of normalizers? (c) Do different tasks and datasets prefer different normalizers? We hope SN will help ease the usage and understand the normalization techniques in deep learning. The code of SN has been released at https://github.com/switchablenorms.
Although skeleton-based action recognition has achieved great success in recent years, most of the existing methods may suffer from a large model size and slow execution speed. To alleviate this issue, we analyze skeleton sequence properties to propose a Double-feature Double-motion Network (DD-Net) for skeleton-based action recognition. By using a lightweight network structure (i.e., 0.15 million parameters), DD-Net can reach a super fast speed, as 3,500 FPS on one GPU, or, 2,000 FPS on one CPU. By employing robust features, DD-Net achieves the state-of-the-art performance on our experimental datasets: SHREC (i.e., hand actions) and JHMDB (i.e., body actions). Our code will be released with this paper later.
Cosmic ray (CR) identification and replacement are critical components of imaging and spectroscopic reduction pipelines involving solid-state detectors. We present deepCR, a deep learning based framework for CR identification and subsequent image inpainting based on the predicted CR mask. To demonstrate the effectiveness of this framework, we train and evaluate models on Hubble Space Telescope ACS/WFC images of sparse extragalactic fields, globular clusters, and resolved galaxies. We demonstrate that at a false positive rate of 0.5%, deepCR achieves close to 100% detection rates in both extragalactic and globular cluster fields, and 91% in resolved galaxy fields, which is a significant improvement over the current state-of-the-art method LACosmic. Compared to a multicore CPU implementation of LACosmic, deepCR CR mask predictions run up to 6.5 times faster on CPU and 90 times faster on a single GPU. For image inpainting, the mean squared errors of deepCR predictions are 20 times lower in globular cluster fields, 5 times lower in resolved galaxy fields, and 2.5 times lower in extragalactic fields, compared to the best performing non-neural technique tested. We present our framework and the trained models as an open-source Python project, with a simple-to-use API. To facilitate reproducibility of the results we also provide a benchmarking codebase.
One-Shot Neural architecture search (NAS) attracts broad attention recently due to its capacity to reduce the computational hours through weight sharing. However, extensive experiments on several recent works show that there is no positive correlation between the validation accuracy with inherited weights from the supernet and the test accuracy after re-training for One-Shot NAS. Different from devising a controller to find the best performing architecture with inherited weights, this paper focuses on how to sample architectures to train the supernet to make it more predictive. A single-path supernet is adopted, where only a small part of weights are optimized in each step, to reduce the memory demand greatly. Furthermore, we abandon devising complicated reward based architecture sampling controller, and sample architectures to train supernet based on novelty search. An efficient novelty search method for NAS is devised in this paper, and extensive experiments demonstrate the effectiveness and efficiency of our novelty search based architecture sampling method. The best architecture obtained by our algorithm with the same search space achieves the state-of-the-art test error rate of 2.51\% on CIFAR-10 with only 7.5 hours search time in a single GPU, and a validation perplexity of 60.02 and a test perplexity of 57.36 on PTB. We also transfer these search cell structures to larger datasets ImageNet and WikiText-2, respectively.
We introduce CuLE (CUDA Learning Environment), a CUDA port of the Atari Learning Environment (ALE) which is used for the development of deep reinforcement algorithms. CuLE overcomes many limitations of existing CPU-based emulators and scales naturally to multiple GPUs. It leverages GPU parallelization to run thousands of games simultaneously and it renders frames directly on the GPU, to avoid the bottleneck arising from the limited CPU-GPU communication bandwidth. CuLE generates up to 155M frames per hour on a single GPU, a finding previously achieved only through a cluster of CPUs. Beyond highlighting the differences between CPU and GPU emulators in the context of reinforcement learning, we show how to leverage the high throughput of CuLE by effective batching of the training data, and show accelerated convergence for A2C+V-trace. CuLE is available at https://github.com/NVLabs/cule .
The term Neural Architecture Search (NAS) refers to the automatic optimization of network architectures for a new, previously unknown task. Since testing an architecture is computationally very expensive, many optimizers need days or even weeks to find suitable architectures. However, this search time can be significantly reduced if knowledge from previous searches on different tasks is reused. In this work, we propose a generally applicable framework that introduces only minor changes to existing optimizers to leverage this feature. As an example, we select an existing optimizer and demonstrate the complexity of the integration of the framework as well as its impact. In experiments on CIFAR-10 and CIFAR-100, we observe a reduction in the search time from 200 to only 6 GPU days, a speed up by a factor of 33. In addition, we observe new records of 1.99 and 14.06 for NAS optimizers on the CIFAR benchmarks, respectively. In a separate study, we analyze the impact of the amount of source and target data. Empirically, we demonstrate that the proposed framework generally gives better results and, in the worst case, is just as good as the unmodified optimizer.