We present an optimized weighted finite-state transducer (WFST) decoder capable of online streaming and offline batch processing of audio using Graphics Processing Units (GPUs). The decoder is efficient in memory utilization, input/output (I/O) bandwidth, and uses a novel Viterbi implementation designed to maximize parallelism. The reduced memory footprint allows the decoder to process significantly larger graphs than previously possible, while optimizing I/O increases the number of simultaneous streams supported. GPU preprocessing of lattice segments enables intermediate lattice results to be returned to the requestor during streaming inference. Collectively, the proposed algorithm yields up to a 240x speedup over single core CPU decoding, and up to 40x faster decoding than the current state-of-the-art GPU decoder, while returning equivalent results. This decoder design enables deployment of production-grade ASR models on a large spectrum of systems, ranging from large data center servers to low-power edge devices.
Neural Architecture Search (NAS) represents a class of methods to generate the optimal neural network architecture and typically iterate over candidate architectures till convergence over some particular metric like validation loss. They are constrained by the available computation resources, especially in enterprise environments. In this paper, we propose a new approach for NAS, called NASIB, which adapts and attunes to the computation resources (budget) available by varying the exploration vs. exploitation trade-off. We reduce the expert bias by searching over an augmented search space induced by Superkernels. The proposed method can provide the architecture search useful for different computation resources and different domains beyond image classification of natural images where we lack bespoke architecture motifs and domain expertise. We show, on CIFAR10, that itis possible to search over a space that comprises of 12x more candidate operations than the traditional prior art in just 1.5 GPU days, while reaching close to state of the art accuracy. While our method searches over an exponentially larger search space, it could lead to novel architectures that require lesser domain expertise, compared to the majority of the existing methods.
Training and deploying deep learning models in real-world applications require processing large amounts of data. This is a challenging task when the amount of data grows to a hundred terabytes, or even, petabyte-scale. We introduce a hybrid distributed cloud framework with a unified view to multiple clouds and an on-premise infrastructure for processing tasks using both CPU and GPU compute instances at scale. The system implements a distributed file system and failure-tolerant task processing scheduler, independent of the language and Deep Learning framework used. It allows to utilize unstable cheap resources on the cloud to significantly reduce costs. We demonstrate the scalability of the framework on running pre-processing, distributed training, hyperparameter search and large-scale inference tasks utilizing 10,000 CPU cores and 300 GPU instances with the overall processing power of 30 petaflops.
Previous works (Donahue et al., 2018a; Engel et al., 2019a) have found that generating coherent raw audio waveforms with GANs is challenging. In this paper, we show that it is possible to train GANs reliably to generate high quality coherent waveforms by introducing a set of architectural changes and simple training techniques. Subjective evaluation metric (Mean Opinion Score, or MOS) shows the effectiveness of the proposed approach for high quality mel-spectrogram inversion. To establish the generality of the proposed techniques, we show qualitative results of our model in speech synthesis, music domain translation and unconditional music synthesis. We evaluate the various components of the model through ablation studies and suggest a set of guidelines to design general purpose discriminators and generators for conditional sequence synthesis tasks. Our model is non-autoregressive, fully convolutional, with significantly fewer parameters than competing models and generalizes to unseen speakers for mel-spectrogram inversion. Our pytorch implementation runs at more than 100x faster than realtime on GTX 1080Ti GPU and more than 2x faster than real-time on CPU, without any hardware specific optimization tricks.
Video understanding is a challenging problem with great impact on the abilities of autonomous agents working in the real-world. Yet, solutions so far have been computationally intensive, with the fastest algorithms running for more than half a second per video snippet on powerful GPUs. We propose a novel idea on video architecture learning - Tiny Video Networks - which automatically designs highly efficient models for video understanding. The tiny video models run with competitive performance for as low as 37 milliseconds per video on a CPU and 10 milliseconds on a standard GPU.
Moving object detection is a critical task for autonomous vehicles. As dynamic objects represent higher collision risk than static ones, our own ego-trajectories have to be planned attending to the future states of the moving elements of the scene. Motion can be perceived using temporal information such as optical flow. Conventional optical flow computation is based on camera sensors only, which makes it prone to failure in conditions with low illumination. On the other hand, LiDAR sensors are independent of illumination, as they measure the time-of-flight of their own emitted lasers. In this work, we propose a robust and real-time CNN architecture for Moving Object Detection (MOD) under low-light conditions by capturing motion information from both camera and LiDAR sensors. We demonstrate the impact of our algorithm on KITTI dataset where we simulate a low-light environment creating a novel dataset "Dark KITTI". We obtain a 10.1% relative improvement on Dark-KITTI, and a 4.25% improvement on standard KITTI relative to our baselines. The proposed algorithm runs at 18 fps on a standard desktop GPU using $256\times1224$ resolution images.
The effectiveness of deep neural networks (DNN) in vision, speech, and language processing has prompted a tremendous demand for energy-efficient high-performance DNN inference systems. Due to the increasing memory intensity of most DNN workloads, main memory can dominate the system's energy consumption and stall time. One effective way to reduce the energy consumption and increase the performance of DNN inference systems is by using approximate memory, which operates with reduced supply voltage and reduced access latency parameters that violate standard specifications. Using approximate memory reduces reliability, leading to higher bit error rates. Fortunately, neural networks have an intrinsic capacity to tolerate increased bit errors. This can enable energy-efficient and high-performance neural network inference using approximate DRAM devices. Based on this observation, we propose EDEN, a general framework that reduces DNN energy consumption and DNN evaluation latency by using approximate DRAM devices, while strictly meeting a user-specified target DNN accuracy. EDEN relies on two key ideas: 1) retraining the DNN for a target approximate DRAM device to increase the DNN's error tolerance, and 2) efficient mapping of the error tolerance of each individual DNN data type to a corresponding approximate DRAM partition in a way that meets the user-specified DNN accuracy requirements. We evaluate EDEN on multi-core CPUs, GPUs, and DNN accelerators with error models obtained from real approximate DRAM devices. For a target accuracy within 1% of the original DNN, our results show that EDEN enables 1) an average DRAM energy reduction of 21%, 37%, 31%, and 32% in CPU, GPU, and two DNN accelerator architectures, respectively, across a variety of DNNs, and 2) an average (maximum) speedup of 8% (17%) and 2.7% (5.5%) in CPU and GPU architectures, respectively, when evaluating latency-bound DNNs.
We designed and implemented a deep learning based RF signal classifier on the Field Programmable Gate Array (FPGA) of an embedded software-defined radio platform, DeepRadio, that classifies the signals received through the RF front end to different modulation types in real time and with low power. This classifier implementation successfully captures complex characteristics of wireless signals to serve critical applications in wireless security and communications systems such as identifying spoofing signals in signal authentication systems, detecting target emitters and jammers in electronic warfare (EW) applications, discriminating primary and secondary users in cognitive radio networks, interference hunting, and adaptive modulation. Empowered by low-power and low-latency embedded computing, the deep neural network runs directly on the FPGA fabric of DeepRadio, while maintaining classifier accuracy close to the software performance. We evaluated the performance when another SDR (USRP) transmits signals with different modulation types at different power levels and DeepRadio receives the signals and classifies them in real time on its FPGA. A smartphone with a mobile app is connected to DeepRadio to initiate the experiment and visualize the classification results. With real radio transmissions over the air, we show that the classifier implemented on DeepRadio achieves high accuracy with low latency (microsecond per sample) and low energy consumption (microJoule per sample), and this performance is not matched by other embedded platforms such as embedded graphics processing unit (GPU).
Smart and agile drones are fast becoming ubiquitous at the edge of the cloud. The usage of these drones are constrained by their limited power and compute capability. In this paper, we present a Transfer Learning (TL) based approach to reduce on-board computation required to train a deep neural network for autonomous navigation via Deep Reinforcement Learning for a target algorithmic performance. A library of 3D realistic meta-environments is manually designed using Unreal Gaming Engine and the network is trained end-to-end. These trained meta-weights are then used as initializers to the network in a test environment and fine-tuned for the last few fully connected layers. Variation in drone dynamics and environmental characteristics is carried out to show robustness of the approach. Using NVIDIA GPU profiler it was shown that the energy consumption and training latency is reduced by 3.7x and 1.8x respectively without significant degradation in the performance in terms of average distance traveled before crash i.e. Mean Safe Flight (MSF). The approach is also tested on a real environment using DJI Tello drone and similar results were reported.
We present the design and optimization of a linear solver on General Purpose GPUs for the efficient and high-throughput evaluation of the marginalized graph kernel between pairs of labeled graphs. The solver implements a preconditioned conjugate gradient (PCG) method to compute the solution to a generalized Laplacian equation associated with the tensor product of two graphs. To cope with the gap between the instruction throughput and the memory bandwidth of current generation GPUs, our solver forms the tensor product linear system on-the-fly without storing it in memory when performing matrix-vector dot product operations in PCG. Such on-the-fly computation is accomplished by using threads in a warp to cooperatively stream the adjacency and edge label matrices of individual graphs by small square matrix blocks called tiles, which are then staged in registers and the shared memory for later reuse. Warps across a thread block can further share tiles via the shared memory to increase data reuse. We exploit the sparsity of the graphs hierarchically by storing only non-empty tiles using a coordinate format and nonzero elements within each tile using bitmaps. Besides, we propose a new partition-based reordering algorithm for aggregating nonzero elements of the graphs into fewer but denser tiles to improve the efficiency of the sparse format. We carry out extensive theoretical analyses on the graph tensor product primitives for tiles of various density and evaluate their performance on synthetic and real-world datasets. Our solver delivers three to four orders of magnitude speedup over existing CPU-based solvers such as GraKeL and GraphKernels. The capability of the solver enables kernel-based learning tasks at unprecedented scales.
The automatic detection and tracking of general objects (like persons, animals or cars), text and logos in a video is crucial for many video understanding tasks, and usually real-time processing as required. We propose OmniTrack, an efficient and robust algorithm which is able to automatically detect and track objects, text as well as brand logos in real-time. It combines a powerful deep learning based object detector (YoloV3) with high-quality optical flow methods. Based on the reference YoloV3 C++ implementation, we did some important performance optimizations which will be described. The major steps in the training procedure for the combined detector for text and logo will be presented. We will describe then the OmniTrack algorithm, consisting of the phases preprocessing, feature calculation, prediction, matching and update. Several performance optimizations have been implemented there as well, like doing the object detection and optical flow calculation asynchronously. Experiments show that the proposed algorithm runs in real-time for standard definition ($720x576$) video on a PC with a Quadro RTX 5000 GPU.
Modern deep learning models have been exploited in various domains, including computer vision (CV), natural language processing (NLP), search and recommendation. In practical AI clusters, workloads training these models are run using software frameworks such as TensorFlow, Caffe, PyTorch and CNTK. One critical issue for efficiently operating practical AI clouds, is to characterize the computing and data transfer demands of these workloads, and more importantly, the training performance given the underlying software framework and hardware configurations. In this paper, we characterize deep learning training workloads from Platform of Artificial Intelligence (PAI) in Alibaba. We establish an analytical framework to investigate detailed execution time breakdown of various workloads using different training architectures, to identify performance bottleneck. Results show that weight/gradient communication during training takes almost 62% of the total execution time among all our workloads on average. The computation part, involving both GPU computing and memory access, are not the biggest bottleneck based on collective behavior of the workloads. We further evaluate attainable performance of the workloads on various potential software/hardware mappings, and explore implications on software architecture selection and hardware configurations. We identify that 60% of PS/Worker workloads can be potentially sped up when ported to the AllReduce architecture exploiting the high-speed NVLink for GPU interconnect, and on average 1.7X speedup can be achieved when Ethernet bandwidth is upgraded from 25 Gbps to 100 Gbps.
The 3D ultrasound (US) entrance inspires a multitude of automated prenatal examinations. However, studies about the structuralized description of the whole fetus in 3D US are still rare. In this paper, we propose to estimate the 3D pose of fetus in US volumes to facilitate its quantitative analyses in global and local scales. Given the great challenges in 3D US, including the high volume dimension, poor image quality, symmetric ambiguity in anatomical structures and large variations of fetal pose, our contribution is three-fold. (i) This is the first work about 3D pose estimation of fetus in the literature. We aim to extract the skeleton of whole fetus and assign different segments/joints with correct torso/limb labels. (ii) We propose a self-supervised learning (SSL) framework to finetune the deep network to form visually plausible pose predictions. Specifically, we leverage the landmark-based registration to effectively encode case-adaptive anatomical priors and generate evolving label proxy for supervision. (iii) To enable our 3D network perceive better contextual cues with higher resolution input under limited computing resource, we further adopt the gradient check-pointing (GCP) strategy to save GPU memory and improve the prediction. Extensively validated on a large 3D US dataset, our method tackles varying fetal poses and achieves promising results. 3D pose estimation of fetus has potentials in serving as a map to provide navigation for many advanced studies.
Histopathological image analysis is an essential process for the discovery of diseases such as cancer. However, it is challenging to train CNN on whole slide images (WSIs) of gigapixel resolution considering the available memory capacity. Most of the previous works divide high resolution WSIs into small image patches and separately input them into the model to classify it as a tumor or a normal tissue. However, patch-based classification uses only patch-scale local information but ignores the relationship between neighboring patches. If we consider the relationship of neighboring patches and global features, we can improve the classification performance. In this paper, we propose a new model structure combining the patch-based classification model and whole slide-scale segmentation model in order to improve the prediction performance of automatic pathological diagnosis. We extract patch features from the classification model and input them into the segmentation model to obtain a whole slide tumor probability heatmap. The classification model considers patch-scale local features, and the segmentation model can take global information into account. We also propose a new optimization method that retains gradient information and trains the model partially for end-to-end learning with limited GPU memory capacity. We apply our method to the tumor/normal prediction on WSIs and the classification performance is improved compared with the conventional patch-based method.
Conventional neural architecture search (NAS) approaches are based on reinforcement learning or evolutionary strategy, which take more than 3000 GPU hours to find a good model on CIFAR-10. We propose an efficient NAS approach learning to search by gradient descent. Our approach represents the search space as a directed acyclic graph (DAG). This DAG contains billions of sub-graphs, each of which indicates a kind of neural architecture. To avoid traversing all the possibilities of the sub-graphs, we develop a differentiable sampler over the DAG. This sampler is learnable and optimized by the validation loss after training the sampled architecture. In this way, our approach can be trained in an end-to-end fashion by gradient descent, named Gradient-based search using Differentiable Architecture Sampler (GDAS). In experiments, we can finish one searching procedure in four GPU hours on CIFAR-10, and the discovered model obtains a test error of 2.82\% with only 2.5M parameters, which is on par with the state-of-the-art. Code is publicly available on GitHub: https://github.com/D-X-Y/NAS-Projects.
Traditional sequence-to-sequence (seq2seq) models and other variations of the attention-mechanism such as hierarchical attention have been applied to the text summarization problem. Though there is a hierarchy in the way humans use language by forming paragraphs from sentences and sentences from words, hierarchical models have usually not worked that much better than their traditional seq2seq counterparts. This effect is mainly because either the hierarchical attention mechanisms are too sparse using hard attention or noisy using soft attention. In this paper, we propose a method based on extracting the highlights of a document; a key concept that is conveyed in a few sentences. In a typical text summarization dataset consisting of documents that are 800 tokens in length (average), capturing long-term dependencies is very important, e.g., the last sentence can be grouped with the first sentence of a document to form a summary. LSTMs (Long Short-Term Memory) proved useful for machine translation. However, they often fail to capture long-term dependencies while modeling long sequences. To address these issues, we have adapted Neural Semantic Encoders (NSE) to text summarization, a class of memory-augmented neural networks by improving its functionalities and proposed a novel hierarchical NSE that outperforms similar previous models significantly. The quality of summarization was improved by augmenting linguistic factors, namely lemma, and Part-of-Speech (PoS) tags, to each word in the dataset for improved vocabulary coverage and generalization. The hierarchical NSE model on factored dataset outperformed the state-of-the-art by nearly 4 ROUGE points. We further designed and used the first GPU-based self-critical Reinforcement Learning model.
With the emergence of onboard vision processing for areas such as the internet of things (IoT), edge computing and autonomous robots, there is increasing demand for computationally efficient convolutional neural network (CNN) models to perform real-time object detection on resource constraints hardware devices. Tiny-YOLO is generally considered as one of the faster object detectors for low-end devices and is the basis for our work. Our experiments on this network have shown that Tiny-YOLO can achieve 0.14 frames per second(FPS) on the Raspberry Pi 3 B, which is too slow for soccer playing autonomous humanoid robots detecting goal and ball objects. In this paper we propose an adaptation to the YOLO CNN model named xYOLO, that can achieve object detection at a speed of 9.66 FPS on the Raspberry Pi 3 B. This is achieved by trading an acceptable amount of accuracy, making the network approximately 70 times faster than Tiny-YOLO. Greater inference speed-ups were also achieved on a desktop CPU and GPU. Additionally we contribute an annotated Darknet dataset for goal and ball detection.
Designing energy-efficient networks is of critical importance for enabling state-of-the-art deep learning in mobile and edge settings where the computation and energy budgets are highly limited. Recently, Liu et al. (2019) framed the search of efficient neural architectures into a continuous splitting process: it iteratively splits existing neurons into multiple off-springs to achieve progressive loss minimization, thus finding novel architectures by gradually growing the neural network. However, this method was not specifically tailored for designing energy-efficient networks, and is computationally expensive on large-scale benchmarks. In this work, we substantially improve Liu et al. (2019) in two significant ways: 1) we incorporate the energy cost of splitting different neurons to better guide the splitting process, thereby discovering more energy-efficient network architectures; 2) we substantially speed up the splitting process of Liu et al. (2019), which requires expensive eigen-decomposition, by proposing a highly scalable Rayleigh-quotient stochastic gradient algorithm. Our fast algorithm allows us to reduce the computational cost of splitting to the same level of typical back-propagation updates and enables efficient implementation on GPU. Extensive empirical results show that our method can train highly accurate and energy-efficient networks on challenging datasets such as ImageNet, improving a variety of baselines, including the pruning-based methods and expert-designed architectures.
As opportunities for AI-assisted healthcare grow steadily, model deployment faces challenges due to the specific characteristics of the industry. The configuration choice for a production device can impact model performance while influencing operational costs. Moreover, in healthcare some situations might require fast, but not real time, inference. We study different configurations and conduct a cost-performance analysis to determine the optimized hardware for the deployment of a model subject to healthcare domain constraints. We observe that a naive performance comparison may not lead to an optimal configuration selection. In fact, given realistic domain constraints, CPU execution might be preferable to GPU accelerators. Hence, defining beforehand precise expectations for model deployment is crucial.
There is a growing interest in automated neural architecture search (NAS). To improve the efficiency of NAS, previous approaches adopt weight sharing method to force all models share the same set of weights. However, it has been observed that a model performing better with shared weights does not necessarily perform better when trained alone. In this paper, we analyse existing weight sharing one-shot NAS approaches from a Bayesian point of view and identify the posterior fading problem, which compromises the effectiveness of shared weights. To alleviate this problem, we present a practical approach to guide the parameter posterior towards its true distribution. Moreover, a hard latency constraint is introduced during the search so that the desired latency can be achieved. The resulted method, namely Posterior Convergent NAS (PC-NAS), achieves state-of-the-art performance under standard GPU latency constraint on ImageNet. In our small search space, our model PC-NAS-S attains 76.8 % top-1 accuracy, 2.1% higher than MobileNetV2 (1.4x) with the same latency. When adopted to the large search space, PC-NAS-L achieves 78.1 % top-1 accuracy within 11ms. The discovered architecture also transfers well to other computer vision applications such as object detection and person re-identification.