Deep neural networks are widely used in personalized recommendation systems. Unlike regular DNN inference workloads, recommendation inference is memory-bound due to the many random memory accesses needed to lookup the embedding tables. The inference is also heavily constrained in terms of latency because producing a recommendation for a user must be done in about tens of milliseconds. In this paper, we propose MicroRec, a high-performance inference engine for recommendation systems. MicroRec accelerates recommendation inference by (1) redesigning the data structures involved in the embeddings to reduce the number of lookups needed and (2) taking advantage of the availability of High-Bandwidth Memory (HBM) in FPGA accelerators to tackle the latency by enabling parallel lookups. We have implemented the resulting design on an FPGA board including the embedding lookup step as well as the complete inference process. Compared to the optimized CPU baseline (16 vCPU, AVX2-enabled), MicroRec achieves 13.8~14.7x speedup on embedding lookup alone and 2.5$~5.4x speedup for the entire recommendation inference in terms of throughput. As for latency, CPU-based engines needs milliseconds for inferring a recommendation while MicroRec only takes microseconds, a significant advantage in real-time recommendation systems.
As an essential processing step in computer vision applications, image resizing or scaling, more specifically downsampling, has to be applied before feeding a normally large image into a convolutional neural network (CNN) model because CNN models typically take small fixed-size images as inputs. However, image scaling functions could be adversarially abused to perform a newly revealed attack called image-scaling attack, which can affect a wide range of computer vision applications building upon image-scaling functions. This work presents an image-scaling attack detection framework, termed as Decamouflage. Decamouflage consists of three independent detection methods: (1) rescaling, (2) filtering/pooling, and (3) steganalysis. While each of these three methods is efficient standalone, they can work in an ensemble manner not only to improve the detection accuracy but also to harden potential adaptive attacks. Decamouflage has a pre-determined detection threshold that is generic. More precisely, as we have validated, the threshold determined from one dataset is also applicable to other different datasets. Extensive experiments show that Decamouflage achieves detection accuracy of 99.9\% and 99.8\% in the white-box (with the knowledge of attack algorithms) and the black-box (without the knowledge of attack algorithms) settings, respectively. To corroborate the efficiency of Decamouflage, we have also measured its run-time overhead on a personal PC with an i5 CPU and found that Decamouflage can detect image-scaling attacks in milliseconds. Overall, Decamouflage can accurately detect image scaling attacks in both white-box and black-box settings with acceptable run-time overhead.
CPU simulators are useful tools for modeling CPU execution behavior. However, they suffer from inaccuracies due to the cost and complexity of setting their fine-grained parameters, such as the latencies of individual instructions. This complexity arises from the expertise required to design benchmarks and measurement frameworks that can precisely measure the values of parameters at such fine granularity. In some cases, these parameters do not necessarily have a physical realization and are therefore fundamentally approximate, or even unmeasurable. In this paper we present DiffTune, a system for learning the parameters of x86 basic block CPU simulators from coarse-grained end-to-end measurements. Given a simulator, DiffTune learns its parameters by first replacing the original simulator with a differentiable surrogate, another function that approximates the original function; by making the surrogate differentiable, DiffTune is then able to apply gradient-based optimization techniques even when the original function is non-differentiable, such as is the case with CPU simulators. With this differentiable surrogate, DiffTune then applies gradient-based optimization to produce values of the simulator's parameters that minimize the simulator's error on a dataset of ground truth end-to-end performance measurements. Finally, the learned parameters are plugged back into the original simulator. DiffTune is able to automatically learn the entire set of microarchitecture-specific parameters within the Intel x86 simulation model of llvm-mca, a basic block CPU simulator based on LLVM's instruction scheduling model. DiffTune's learned parameters lead llvm-mca to an average error that not only matches but lowers that of its original, expert-provided parameter values.
We present 'Pix2Prof', a deep learning model that can eliminate any manual steps taken when extracting galaxy profiles. We argue that a galaxy profile of any sort is conceptually similar to a natural language image caption. This idea allows us to leverage image captioning methods from the field of natural language processing, and so we design Pix2Prof as a float sequence 'captioning' model suitable for galaxy profile inference. We demonstrate the technique by approximating a galaxy surface brightness (SB) profile fitting method that contains several manual steps. Pix2Prof processes $\sim$1 image per second on an Intel Xeon E5 2650 v3 CPU, improving on the speed of the manual interactive method by more than two orders of magnitude. Crucially, Pix2Prof requires no manual interaction, and since galaxy profile estimation is an embarrassingly parallel problem, we can further increase the throughput by running many Pix2Prof instances simultaneously. In perspective, Pix2Prof would take under an hour to infer profiles for $10^5$ galaxies on a single NVIDIA DGX-2 system. A single human expert would take approximately two years to complete the same task. Automated methodology such as this will accelerate the analysis of the next generation of large area sky surveys expected to yield hundreds of millions of targets. In such instances, all manual approaches -- even those involving a large number of experts -- will be impractical.
We present ELQ, a fast end-to-end entity linking model for questions, which uses a biencoder to jointly perform mention detection and linking in one pass. Evaluated on WebQSP and GraphQuestions with extended annotations that cover multiple entities per question, ELQ outperforms the previous state of the art by a large margin of +12.7% and +19.6% F1, respectively. With a very fast inference time (1.57 examples/s on a single CPU), ELQ can be useful for downstream question answering systems. In a proof-of-concept experiment, we demonstrate that using ELQ significantly improves the downstream QA performance of GraphRetriever (arXiv:1911.03868). Code and data available at https://github.com/facebookresearch/BLINK/tree/master/elq
A lot of deep learning applications are desired to be run on mobile devices. Both accuracy and inference time are meaningful for a lot of them. While the number of FLOPs is usually used as a proxy for neural network latency, it may be not the best choice. In order to obtain a better approximation of latency, research community uses look-up tables of all possible layers for latency calculation for the final prediction of the inference on mobile CPU. It requires only a small number of experiments. Unfortunately, on mobile GPU this method is not applicable in a straight-forward way and shows low precision. In this work, we consider latency approximation on mobile GPU as a data and hardware-specific problem. Our main goal is to construct a convenient latency estimation tool for investigation(LETI) of neural network inference and building robust and accurate latency prediction models for each specific task. To achieve this goal, we build open-source tools which provide a convenient way to conduct massive experiments on different target devices focusing on mobile GPU. After evaluation of the dataset, we learn the regression model on experimental data and use it for future latency prediction and analysis. We experimentally demonstrate the applicability of such an approach on a subset of popular NAS-Benchmark 101 dataset and also evaluate the most popular neural network architectures for two mobile GPUs. As a result, we construct latency prediction model with good precision on the target evaluation subset. We consider LETI as a useful tool for neural architecture search or massive latency evaluation. The project is available at https://github.com/leti-ai
Large Transformer models have achieved state-of-the-art results in neural machine translation and have become standard in the field. In this work, we look for the optimal combination of known techniques to optimize inference speed without sacrificing translation quality. We conduct an empirical study that stacks various approaches and demonstrates that combination of replacing decoder self-attention with simplified recurrent units, adopting a deep encoder and a shallow decoder architecture and multi-head attention pruning can achieve up to 109% and 84% speedup on CPU and GPU respectively and reduce the number of parameters by 25% while maintaining the same translation quality in terms of BLEU.
As machine learning spreads into more and more application areas, micro controllers and low power CPUs are increasingly being used to perform inference with machine learning models. The capability to deploy onto these limited hardware targets is enabling machine learning models to be used across a diverse range of new domains. Optimising the inference process on these targets poses different challenges from either desktop CPU or GPU implementations, where the small amounts of RAM available on these targets sets limits on size of models which can be executed. Analysis of the memory use patterns of eleven machine learning models was performed. Memory load and store patterns were observed using a modified version of the Valgrind debugging tool, identifying memory areas holding values necessary for the calculation as inference progressed. These analyses identified opportunities optimise the memory use of these models by overlapping the input and output buffers of individual tensor operations. Three methods are presented which can calculate the safe overlap of input and output buffers for tensor operations. Ranging from a computationally expensive approach with the ability to operate on compiled layer operations, to a versatile analytical solution which requires access to the original source code of the layer. The diagonal memory optimisation technique is described and shown to achieve memory savings of up to 34.5% when applied to eleven common models. Micro-controller targets are identified where it is only possible to deploy some models if diagonal memory optimisation is used.
BACKGROUND AND OBJECTIVE: The research done in the field of Augmented Reality (AR) for patient positioning in radiation therapy is scarce. We propose an efficient and cost-effective algorithm for tracking the scene and the patient to interactively assist the patient's positioning process by providing visual feedback to the operator. Up to our knowledge, this is the first framework that can be employed for mobile interactive AR to guide patient positioning. METHODS: We propose a point cloud processing method that combined with a fiducial marker-mapper algorithm and the generalized ICP algorithm tracks the patient and the camera precisely and efficiently only using the CPU unit. The alignment between the 3D reference model and body marker map is calculated employing an efficient body reconstruction algorithm. RESULTS: Our quantitative evaluation shows that the proposed method achieves a translational and rotational error of 4.17 mm/0.82 deg at 9 fps. Furthermore, the qualitative results demonstrate the usefulness of our algorithm in patient positioning on different human subjects. CONCLUSION: Since our algorithm achieves a relatively high frame rate and accuracy employing a regular laptop (without the usage of a dedicated GPU), it is a very cost-effective AR-based patient positioning method. It also opens the way for other researchers by introducing a framework that could be improved upon for better mobile interactive AR patient positioning solutions in the future.
Out-of-home audience measurement aims to count and characterize the people exposed to advertising content in the physical world. While audience measurement solutions based on computer vision are of increasing interest, no commonly accepted benchmark exists to evaluate and compare their performance. In this paper, we propose the first benchmark for digital out-of-home audience measurement that evaluates the vision-based tasks of audience localization and counting, and audience demographics. The benchmark is composed of a novel, dataset captured at multiple locations and a set of performance measures. Using the benchmark, we present an in-depth comparison of eight open-source algorithms on four hardware platforms with GPU and CPU-optimized inferences and of two commercial off-the-shelf solutions for localization, count, age, and gender estimation. This benchmark and related open-source codes are available at http://ava.eecs.qmul.ac.uk.
We show in this work that memory intensive computations can result in severe performance problems due to off-chip memory access and CPU-GPU context switch overheads in a wide range of deep learning models. For this problem, current just-in-time (JIT) kernel fusion and code generation techniques have limitations, such as rough fusion plan exploration strategies and limited code generation ability. We propose FusionStitching, a deep learning compiler capable of fusing memory intensive operators, with varied data dependencies and non-homogeneous parallelism, into large GPU kernels to reduce global memory access and context switch overhead automatically. FusionStitching widens the range of operation combinations that fusion can target beyond previous JIT works by introducing data reuse of intermediate values. It explores large fusion spaces to decide optimal fusion plans with considerations of memory access costs, kernel calls and resource usage constraints. FusionStitching tunes the optimal stitching scheme with a domain-specific cost model efficiently. Experimental results show that FusionStitching can reach up to 2.21x speedup compared to state-of-the-art, with 1.45x on average. Besides these experimental results, we integrated our approach into a compiler product and deployed it onto a production cluster for AI workloads with thousands of GPUs. The system has been in operation for more than 4 months and saves 7,000 GPU hours on average for approximately 30,000 tasks per month.
Cloud data center capacities are over-provisioned to handle demand peaks and hardware failures which leads to low resources' utilization. One way to improve resource utilization and thus reduce the total cost of ownership is to offer unused resources (referred to as ephemeral resources) at a lower price. However, reselling resources needs to meet the expectations of its customers in terms of Quality of Service. The goal is so to maximize the amount of reclaimed resources while avoiding SLA penalties. To achieve that, cloud providers have to estimate their future utilization to provide availability guarantees. The prediction should consider a safety margin for resources to react to unpredictable workloads. The challenge is to find the safety margin that provides the best trade-off between the amount of resources to reclaim and the risk of SLA violations. Most state-of-the-art solutions consider a fixed safety margin for all types of metrics (e.g., CPU, RAM). However, a unique fixed margin does not consider various workloads variations over time which may lead to SLA violations or/and poor utilization. In order to tackle these challenges, we propose ReLeaSER, a Reinforcement Learning strategy for optimizing the ephemeral resources' utilization in the cloud. ReLeaSER dynamically tunes the safety margin at the host-level for each resource metric. The strategy learns from past prediction errors (that caused SLA violations). Our solution reduces significantly the SLA violation penalties on average by 2.7x and up to 3.4x. It also improves considerably the CPs' potential savings by 27.6% on average and up to 43.6%.
Existing methods of 3D dense face alignment mainly concentrate on accuracy, thus limiting the scope of their practical applications. In this paper, we propose a novel regression framework named 3DDFA-V2 which makes a balance among speed, accuracy and stability. Firstly, on the basis of a lightweight backbone, we propose a meta-joint optimization strategy to dynamically regress a small set of 3DMM parameters, which greatly enhances speed and accuracy simultaneously. To further improve the stability on videos, we present a virtual synthesis method to transform one still image to a short-video which incorporates in-plane and out-of-plane face moving. On the premise of high accuracy and stability, 3DDFA-V2 runs at over 50fps on a single CPU core and outperforms other state-of-the-art heavy models simultaneously. Experiments on several challenging datasets validate the efficiency of our method. Pre-trained models and code are available at https://github.com/cleardusk/3DDFA_V2.
In this work, we leverage ensemble learning as a tool for the creation of faster, smaller, and more accurate deep learning models. We demonstrate that we can jointly optimize for accuracy, inference time, and the number of parameters by combining DNN classifiers. To achieve this, we combine multiple ensemble strategies: bagging, boosting, and an ordered chain of classifiers. To reduce the number of DNN ensemble evaluations during the search, we propose EARN, an evolutionary approach that optimizes the ensemble according to three objectives regarding the constraints specified by the user. We run EARN on 10 image classification datasets with an initial pool of 32 state-of-the-art DCNN on both CPU and GPU platforms, and we generate models with speedups up to $7.60\times$, reductions of parameters by $10\times$, or increases in accuracy up to $6.01\%$ regarding the best DNN in the pool. In addition, our method generates models that are $5.6\times$ faster than the state-of-the-art methods for automatic model generation.
Pre-trained large-scale language models have increasingly demonstrated high accuracy on many natural language processing (NLP) tasks. However, the limited weight storage and computational speed on hardware platforms have impeded the popularity of pre-trained models, especially in the era of edge computing. In this paper, we seek to find the best model structure of BERT for a given computation size to match specific devices. We propose the first compiler-aware neural architecture optimization framework. Our framework can guarantee the identified model to meet both resource and real-time specifications of mobile devices, thus achieving real-time execution of large transformer-based models like BERT variants. We evaluate our model on several NLP tasks, achieving competitive results on well-known benchmarks with lower latency on mobile devices. Specifically, our model is 5.2x faster on CPU and 4.1x faster on GPU with 0.5-2% accuracy loss compared with BERT-base. Our overall framework achieves up to 7.8x speedup compared with TensorFlow-Lite with only minor accuracy loss.
We present PDFFlow, a new software for fast evaluation of parton distribution functions (PDFs) designed for platforms with hardware accelerators. PDFs are essential for the calculation of particle physics observables through Monte Carlo simulation techniques. The evaluation of a generic set of PDFs for quarks and gluon at a given momentum fraction and energy scale requires the implementation of interpolation algorithms as introduced for the first time by the LHAPDF project. PDFFlow extends and implements these interpolation algorithms using Google's TensorFlow library providing the capabilities to perform PDF evaluations taking fully advantage of multi-threading CPU and GPU setups. We benchmark the performance of this library on multiple scenarios relevant for the particle physics community.
Recent researches on robotics have shown significant improvement, spanning from algorithms, mechanics to hardware architectures. Robotics, including manipulators, legged robots, drones, and autonomous vehicles, are now widely applied in diverse scenarios. However, the high computation and data complexity of robotic algorithms pose great challenges to its applications. On the one hand, CPU platform is flexible to handle multiple robotic tasks. GPU platform has higher computational capacities and easy-touse development frameworks, so they have been widely adopted in several applications. On the other hand, FPGA-based robotic accelerators are becoming increasingly competitive alternatives, especially in latency-critical and power-limited scenarios. With specialized designed hardware logic and algorithm kernels, FPGA-based accelerators can surpass CPU and GPU in performance and energy efficiency. In this paper, we give an overview of previous work on FPGA-based robotic accelerators covering different stages of the robotic system pipeline. An analysis of software and hardware optimization techniques and main technical issues is presented, along with some commercial and space applications, to serve as a guide for future work.
The rapid development and wide utilization of object detection techniques have aroused attention on both accuracy and speed of object detectors. However, the current state-of-the-art object detection works are either accuracy-oriented using a large model but leading to high latency or speed-oriented using a lightweight model but sacrificing accuracy. In this work, we propose YOLObile framework, a real-time object detection on mobile devices via compression-compilation co-design. A novel block-punched pruning scheme is proposed for any kernel size. To improve computational efficiency on mobile devices, a GPU-CPU collaborative scheme is adopted along with advanced compiler-assisted optimizations. Experimental results indicate that our pruning scheme achieves 14$\times$ compression rate of YOLOv4 with 49.0 mAP. Under our YOLObile framework, we achieve 17 FPS inference speed using GPU on Samsung Galaxy S20. By incorporating our proposed GPU-CPU collaborative scheme, the inference speed is increased to 19.1 FPS, and outperforms the original YOLOv4 by 5$\times$ speedup. Source code is at: \url{https://github.com/nightsnack/YOLObile}.
This paper describes the design, implementation, and evaluation of Otak, a system that allows two non-colluding cloud providers to run machine learning (ML) inference without knowing the inputs to inference. Prior work for this problem mostly relies on advanced cryptography such as two-party secure computation (2PC) protocols that provide rigorous guarantees but suffer from high resource overhead. Otak improves efficiency via a new 2PC protocol that (i) tailors recent primitives such as function and homomorphic secret sharing to ML inference, and (ii) uses trusted hardware in a limited capacity to bootstrap the protocol. At the same time, Otak reduces trust assumptions on trusted hardware by running a small code inside the hardware, restricting its use to a preprocessing step, and distributing trust over heterogeneous trusted hardware platforms from different vendors. An implementation and evaluation of Otak demonstrates that its CPU and network overhead converted to a dollar amount is 5.4$-$385$\times$ lower than state-of-the-art 2PC-based works. Besides, Otak's trusted computing base (code inside trusted hardware) is only 1,300 lines of code, which is 14.6$-$29.2$\times$ lower than the code-size in prior trusted hardware-based works.
B-mode ultrasound is commonly used to image musculoskeletal tissues, but one major bottleneck is data interpretation, and analyses of muscle thickness, pennation angle and fascicle length are often still performed manually. In this study we trained deep neural networks (based on U-net) to detect muscle fascicles and aponeuroses using a set of labelled musculoskeletal ultrasound images. We then compared neural network predictions on new, unseen images to those obtained via manual analysis and two existing semi/automated analysis approaches (SMA and Ultratrack). With a GPU, inference time for a single image with the new approach was around 0.7s, compared to 4.6s with a CPU. Our method detects the locations of the superficial and deep aponeuroses, as well as multiple fascicle fragments per image. For single images, the method gave similar results to those produced by a non-trainable automated method (SMA; mean difference in fascicle length: 1.1 mm) or human manual analysis (mean difference: 2.1 mm). Between-method differences in pennation angle were within 1$^\circ$, and mean differences in muscle thickness were less than 0.2 mm. Similarly, for videos, there was strong overlap between the results produced with Ultratrack and our method, with a mean ICC of 0.73, despite the fact that the analysed trials included hundreds of frames. Our method is fully automated and open source, and can estimate fascicle length, pennation angle and muscle thickness from single images or videos, as well as from multiple superficial muscles. We also provide all necessary code and training data for custom model development.