High-resolution dense prediction enables many appealing real-world applications, such as computational photography, autonomous driving, etc. However, the vast computational cost makes deploying state-of-the-art high-resolution dense prediction models on hardware devices difficult. This work presents EfficientViT, a new family of high-resolution vision models with novel multi-scale linear attention. Unlike prior high-resolution dense prediction models that rely on heavy softmax attention, hardware-inefficient large-kernel convolution, or complicated topology structure to obtain good performances, our multi-scale linear attention achieves the global receptive field and multi-scale learning (two desirable features for high-resolution dense prediction) with only lightweight and hardware-efficient operations. As such, EfficientViT delivers remarkable performance gains over previous state-of-the-art models with significant speedup on diverse hardware platforms, including mobile CPU, edge GPU, and cloud GPU. Without performance loss on Cityscapes, our EfficientViT provides up to 13.9$\times$ and 6.2$\times$ GPU latency reduction over SegFormer and SegNeXt, respectively. For super-resolution, EfficientViT delivers up to 6.4x speedup over Restormer while providing 0.11dB gain in PSNR. For Segment Anything, EfficientViT delivers 48.9x higher throughput on A100 GPU while achieving slightly better zero-shot instance segmentation performance on COCO.
Deep Learning (DL) has transformed the automation of a wide range of industries and finds increasing ubiquity in society. The high complexity of DL models and its widespread adoption has led to global energy consumption doubling every 3-4 months. Currently, the relationship between the DL model configuration and energy consumption is not well established. At a general computational energy model level, there is both strong dependency to both the hardware architecture (e.g. generic processors with different configuration of inner components- CPU and GPU, programmable integrated circuits - FPGA), as well as different interacting energy consumption aspects (e.g., data movement, calculation, control). At the DL model level, we need to translate non-linear activation functions and its interaction with data into calculation tasks. Current methods mainly linearize nonlinear DL models to approximate its theoretical FLOPs and MACs as a proxy for energy consumption. Yet, this is inaccurate (est. 93\% accuracy) due to the highly nonlinear nature of many convolutional neural networks (CNNs) for example. In this paper, we develop a bottom-level Transistor Operations (TOs) method to expose the role of non-linear activation functions and neural network structure in energy consumption. We translate a range of feedforward and CNN models into ALU calculation tasks and then TO steps. This is then statistically linked to real energy consumption values via a regression model for different hardware configurations and data sets. We show that our proposed TOs method can achieve a 93.61% - 99.51% precision in predicting its energy consumption.
We propose a novel neural architecture for computer vision -- WaveMix -- that is resource-efficient and yet generalizable and scalable. While using fewer trainable parameters, GPU RAM, and computations, WaveMix networks achieve comparable or better accuracy than the state-of-the-art convolutional neural networks, vision transformers, and token mixers for several tasks. This efficiency can translate to savings in time, cost, and energy. To achieve these gains we used multi-level two-dimensional discrete wavelet transform (2D-DWT) in WaveMix blocks, which has the following advantages: (1) It reorganizes spatial information based on three strong image priors -- scale-invariance, shift-invariance, and sparseness of edges -- (2) in a lossless manner without adding parameters, (3) while also reducing the spatial sizes of feature maps, which reduces the memory and time required for forward and backward passes, and (4) expanding the receptive field faster than convolutions do. The whole architecture is a stack of self-similar and resolution-preserving WaveMix blocks, which allows architectural flexibility for various tasks and levels of resource availability. WaveMix establishes new benchmarks for segmentation on Cityscapes; and for classification on Galaxy 10 DECals, Places-365, five EMNIST datasets, and iNAT-mini and performs competitively on other benchmarks. Our code and trained models are publicly available.
Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware -- accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3$\times$ speedup on GPT-2 (seq. length 1K), and 2.4$\times$ speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).
On-Shelf Availability (OSA) of products in retail stores is a critical business criterion in the fast moving consumer goods and retails sector. When a product is out-of-stock (OOS) and a customer cannot find it on its designed shelf, this motivates the customer to store-switching or buying nothing, which causes fall in future sales and demands. Retailers are employing several approaches to detect empty shelves and ensure high OSA of products; however, such methods are generally ineffective and infeasible since they are either manual, expensive or less accurate. Recently machine learning based solutions have been proposed, but they suffer from high computational cost and low accuracy problem due to lack of large annotated datasets of on-shelf products. Here, we present an elegant approach for designing an end-to-end machine learning (ML) pipeline for real-time empty shelf detection. Considering the strong dependency between the quality of ML models and the quality of data, we focus on the importance of proper data collection, cleaning and correct data annotation before delving into modeling. Since an empty-shelf detection solution should be computationally-efficient for real-time predictions, we explore different run-time optimizations to improve the model performance. Our dataset contains 1000 images, collected and annotated by following well-defined guidelines. Our low-latency model achieves a mean average F1-score of 68.5%, and can process up to 67 images/s on Intel Xeon Gold and up to 860 images/s on an A100 GPU.
We present an efficient approach for Masked Image Modeling (MIM) with hierarchical Vision Transformers (ViTs), allowing the hierarchical ViTs to discard masked patches and operate only on the visible ones. Our approach consists of three key designs. First, for window attention, we propose a Group Window Attention scheme following the Divide-and-Conquer strategy. To mitigate the quadratic complexity of the self-attention w.r.t. the number of patches, group attention encourages a uniform partition that visible patches within each local window of arbitrary size can be grouped with equal size, where masked self-attention is then performed within each group. Second, we further improve the grouping strategy via the Dynamic Programming algorithm to minimize the overall computation cost of the attention on the grouped patches. Third, as for the convolution layers, we convert them to the Sparse Convolution that works seamlessly with the sparse data, i.e., the visible patches in MIM. As a result, MIM can now work on most, if not all, hierarchical ViTs in a green and efficient way. For example, we can train the hierarchical ViTs, e.g., Swin Transformer and Twins Transformer, about 2.7$\times$ faster and reduce the GPU memory usage by 70%, while still enjoying competitive performance on ImageNet classification and the superiority on downstream COCO object detection benchmarks. Code and pre-trained models have been made publicly available at https://github.com/LayneH/GreenMIM.
Sparse active illumination enables precise time-of-flight depth sensing as it maximizes signal-to-noise ratio for low power budgets. However, depth completion is required to produce dense depth maps for 3D perception. We address this task with realistic illumination and sensor resolution constraints by simulating ToF datasets for indoor 3D perception with challenging sparsity levels. We propose a quantized convolutional encoder-decoder network for this task. Our model achieves optimal depth map quality by means of input pre-processing and carefully tuned training with a geometry-preserving loss function. We also achieve low memory footprint for weights and activations by means of mixed precision quantization-at-training techniques. The resulting quantized models are comparable to the state of the art in terms of quality, but they require very low GPU times and achieve up to 14-fold memory size reduction for the weights w.r.t. their floating point counterpart with minimal impact on quality metrics.
Currently, Deep learning and especially Convolutional Neural Networks (CNNs) have become a fundamental computational approach applied in a wide range of domains, including some safety-critical applications (e.g., automotive, robotics, and healthcare equipment). Therefore, the reliability evaluation of those computational systems is mandatory. The reliability evaluation of CNNs is performed by fault injection campaigns at different levels of abstraction, from the application level down to the hardware level. Many works have focused on evaluating the reliability of neural networks in the presence of transient faults. However, the effects of permanent faults have been investigated at the application level, only, e.g., targeting the parameters of the network. This paper intends to propose a framework, resorting to a binary instrumentation tool to perform fault injection campaigns, targeting different components inside the GPU, such as the register files and the functional units. This environment allows for the first time assessing the reliability of CNNs deployed on a GPU considering the presence of permanent faults.
We present a fast, scalable, data-driven approach for solving relaxations of 0-1 integer linear programs. We use a combination of graph neural networks (GNN) and the Lagrange decomposition based algorithm FastDOG (Abbas and Swoboda 2022b). We make the latter differentiable for end-to-end training and use GNNs to predict its algorithmic parameters. This allows to retain the algorithm's theoretical properties including dual feasibility and guaranteed non-decrease in the lower bound while improving it via training. We overcome suboptimal fixed points of the basic solver by additional non-parametric GNN update steps maintaining dual feasibility. For training we use an unsupervised loss. We train on smaller problems and test on larger ones showing strong generalization performance with a GNN comprising only around $10k$ parameters. Our solver achieves significantly faster performance and better dual objectives than its non-learned version, achieving close to optimal objective values of LP relaxations of very large structured prediction problems and on selected combinatorial ones. In particular, we achieve better objective values than specialized approximate solvers for specific problem classes while retaining their efficiency. Our solver has better any-time performance over a large time period compared to a commercial solver. Code available at https://github.com/LPMP/BDD
Deep learning based computed tomography (CT) reconstruction has demonstrated outstanding performance on simulated 2D low-dose CT data. This applies in particular to domain adapted neural networks, which incorporate a handcrafted physics model for CT imaging. Empirical evidence shows that employing such architectures reduces the demand for training data and improves upon generalisation. However, their training requires large computational resources that quickly become prohibitive in 3D helical CT, which is the most common acquisition geometry used for medical imaging. Furthermore, clinical data also comes with other challenges not accounted for in simulations, like errors in flux measurement, resolution mismatch and, most importantly, the absence of the real ground truth. The necessity to have a computationally feasible training combined with the need to address these issues has made it difficult to evaluate deep learning based reconstruction on clinical 3D helical CT. This paper modifies a domain adapted neural network architecture, the Learned Primal-Dual (LPD), so that it can be trained and applied to reconstruction in this setting. We achieve this by splitting the helical trajectory into sections and applying the unrolled LPD iterations to those sections sequentially. To the best of our knowledge, this work is the first to apply an unrolled deep learning architecture for reconstruction on full-sized clinical data, like those in the Low dose CT image and projection data set (LDCT). Moreover, training and testing is done on a single GPU card with 24GB of memory.
Deep learning (DL) shows its prosperity in a wide variety of fields. The development of a DL model is a time-consuming and resource-intensive procedure. Hence, dedicated GPU accelerators have been collectively constructed into a GPU datacenter. An efficient scheduler design for such GPU datacenter is crucially important to reduce the operational cost and improve resource utilization. However, traditional approaches designed for big data or high performance computing workloads can not support DL workloads to fully utilize the GPU resources. Recently, substantial schedulers are proposed to tailor for DL workloads in GPU datacenters. This paper surveys existing research efforts for both training and inference workloads. We primarily present how existing schedulers facilitate the respective workloads from the scheduling objectives and resource consumption features. Finally, we prospect several promising future research directions. More detailed summary with the surveyed paper and code links can be found at our project website: https://github.com/S-Lab-System-Group/Awesome-DL-Scheduling-Papers
The increased interest in Artificial Intelligence (AI) raised the need for highly optimized and sophisticated AI frameworks. Starting with the Lua-based Torch many frameworks have emerged over time, such as Theano, Caffe, Chainer, CNTK, MxNet, PyTorch, DL4J, or TensorFlow. All of these provide a high level scripting API that allows users to easily design neural networks and run these on various kinds of hardware. What the user usually does not see is the high effort put into these frameworks to provide peak execution performance. While mainstream CPUs and GPUs have the "luxury" to have a wide spread user base in the open source community, less mainstream CPU, GPU or accelerator vendors need to put in a high effort to get their hardware supported by these frameworks. This includes not only the development of highly efficient compute libraries such as CUDNN, OneDNN or VEDNN but also supporting an ever growing number of simpler compute operations such as summation and multiplications. Each of these frameworks, nowadays, supports several hundred of unique operations, with tensors of various sizes, shapes and data types, which end up in thousands of compute kernels required for each device type. And the number of operations keeps increasing. That is why NEC Laboratories Europe started developing the SOL AI Optimization project already years ago, to deliver optimal performance to users while keeping the maintenance burden minimal.
Recent research on human pose estimation exploits complex structures to improve performance on benchmark datasets, ignoring the resource overhead and inference speed when the model is actually deployed. In this paper, we lighten the computation cost and parameters of the deconvolution head network in SimpleBaseline and introduce an attention mechanism that utilizes original, inter-level, and intra-level information to intensify the accuracy. Additionally, we propose a novel loss function called heatmap weighting loss, which generates weights for each pixel on the heatmap that makes the model more focused on keypoints. Experiments demonstrate our method achieves a balance between performance, resource volume, and inference speed. Specifically, our method can achieve 65.3 AP score on COCO test-dev, while the inference speed is 55 FPS and 18 FPS on the mobile GPU and CPU, respectively.
We introduce a novel co-design method for autonomous moving agents' shape attributes and locomotion by combining deep reinforcement learning and evolution with user control. Our main inspiration comes from evolution, which has led to wide variability and adaptation in Nature and has the potential to significantly improve design and behavior simultaneously. Our method takes an input agent with optional simple constraints such as leg parts that should not evolve or allowed ranges of changes. It uses physics-based simulation to determine its locomotion and finds a behavior policy for the input design, later used as a baseline for comparison. The agent is then randomly modified within the allowed ranges creating a new generation of several hundred agents. The generation is trained by transferring the previous policy, which significantly speeds up the training. The best-performing agents are selected, and a new generation is formed using their crossover and mutations. The next generations are then trained until satisfactory results are reached. We show a wide variety of evolved agents, and our results show that even with only 10% of changes, the overall performance of the evolved agents improves 50%. If more significant changes to the initial design are allowed, our experiments' performance improves even more to 150%. Contrary to related work, our co-design works on a single GPU and provides satisfactory results by training thousands of agents within one hour.
Random quantum circuits have been utilized in the contexts of quantum supremacy demonstrations, variational quantum algorithms for chemistry and machine learning, and blackhole information. The ability of random circuits to approximate any random unitaries has consequences on their complexity, expressibility, and trainability. To study this property of random circuits, we develop numerical protocols for estimating the frame potential, the distance between a given ensemble and the exact randomness. Our tensor-network-based algorithm has polynomial complexity for shallow circuits and is high-performing using CPU and GPU parallelism. We study 1. local and parallel random circuits to verify the linear growth in complexity as stated by the Brown-Susskind conjecture, and; 2. hardware-efficient ans\"atze to shed light on its expressibility and the barren plateau problem in the context of variational algorithms. Our work shows that large-scale tensor network simulations could provide important hints toward open problems in quantum information science.
We study lossless acceleration for seq2seq generation with a novel decoding algorithm -- Aggressive Decoding. Unlike the previous efforts (e.g., non-autoregressive decoding) speeding up seq2seq generation at the cost of quality loss, our approach aims to yield the identical (or better) generation compared with autoregressive decoding but in a significant speedup, achieved by innovative cooperation of aggressive decoding and verification that are both efficient due to parallel computing. We propose two Aggressive Decoding paradigms for 2 kinds of seq2seq tasks: 1) For the seq2seq tasks whose inputs and outputs are highly similar (e.g., Grammatical Error Correction), we propose Input-guided Aggressive Decoding (IAD) that aggressively copies from the input sentence as drafted decoded tokens to verify in parallel; 2) For other general seq2seq tasks (e.g., Machine Translation), we propose Generalized Aggressive Decoding (GAD) that first employs an additional non-autoregressive decoding model for aggressive decoding and then verifies in parallel in the autoregressive manner. We test Aggressive Decoding on the most popular 6-layer Transformer model on GPU in multiple seq2seq tasks: 1) For IAD, we show that it can introduce a 7x-9x speedup for the Transformer in Grammatical Error Correction and Text Simplification tasks with the identical results as greedy decoding; 2) For GAD, we observe a 3x-5x speedup with the identical or even better quality in two important seq2seq tasks: Machine Translation and Abstractive Summarization. Moreover, Aggressive Decoding can benefit even more from stronger computing devices that are better at parallel computing. Given the lossless quality as well as significant and promising speedup, we believe Aggressive Decoding may potentially evolve into a de facto standard for efficient and lossless seq2seq generation in the near future.
Spiking Neural Networks (SNNs) are gaining significant traction in machine learning tasks where energy-efficiency is of utmost importance. Training such networks using the state-of-the-art back-propagation through time (BPTT) is, however, very time-consuming. Previous work by Shrestha and Orchard [2018] employs an efficient GPU-accelerated back-propagation algorithm called SLAYER, which speeds up training considerably. SLAYER, however, does not take into account the neuron reset mechanism while computing the gradients, which we argue to be the source of numerical instability. To counteract this, SLAYER introduces a gradient scale hyperparameter across layers, which needs manual tuning. In this paper, (i) we modify SLAYER and design an algorithm called EXODUS, that accounts for the neuron reset mechanism and applies the Implicit Function Theorem (IFT) to calculate the correct gradients (equivalent to those computed by BPTT), (ii) we eliminate the need for ad-hoc scaling of gradients, thus, reducing the training complexity tremendously, (iii) we demonstrate, via computer simulations, that EXODUS is numerically stable and achieves a comparable or better performance than SLAYER especially in various tasks with SNNs that rely on temporal features. Our code is available at https://github.com/synsense/sinabs-exodus.
Masked AutoEncoder (MAE) has recently led the trends of visual self-supervision area by an elegant asymmetric encoder-decoder design, which significantly optimizes both the pre-training efficiency and fine-tuning accuracy. Notably, the success of the asymmetric structure relies on the "global" property of Vanilla Vision Transformer (ViT), whose self-attention mechanism reasons over arbitrary subset of discrete image patches. However, it is still unclear how the advanced Pyramid-based ViTs (e.g., PVT, Swin) can be adopted in MAE pre-training as they commonly introduce operators within "local" windows, making it difficult to handle the random sequence of partial vision tokens. In this paper, we propose Uniform Masking (UM), successfully enabling MAE pre-training for Pyramid-based ViTs with locality (termed "UM-MAE" for short). Specifically, UM includes a Uniform Sampling (US) that strictly samples $1$ random patch from each $2 \times 2$ grid, and a Secondary Masking (SM) which randomly masks a portion of (usually $25\%$) the already sampled regions as learnable tokens. US preserves equivalent elements across multiple non-overlapped local windows, resulting in the smooth support for popular Pyramid-based ViTs; whilst SM is designed for better transferable visual representations since US reduces the difficulty of pixel recovery pre-task that hinders the semantic learning. We demonstrate that UM-MAE significantly improves the pre-training efficiency (e.g., it speeds up and reduces the GPU memory by $\sim 2\times$) of Pyramid-based ViTs, but maintains the competitive fine-tuning performance across downstream tasks. For example using HTC++ detector, the pre-trained Swin-Large backbone self-supervised under UM-MAE only in ImageNet-1K can even outperform the one supervised in ImageNet-22K. The codes are available at https://github.com/implus/UM-MAE.
While modern internet services, such as chatbots, search engines, and online advertising, demand the use of large-scale deep neural networks (DNNs), distributed training and inference over heterogeneous computing systems are desired to facilitate these DNN models. Mixture-of-Experts (MoE) is one the most common strategies to lower the cost of training subject to the overall size of models/data through gating and parallelism in a divide-and-conquer fashion. While DeepSpeed has made efforts in carrying out large-scale MoE training over heterogeneous infrastructures, the efficiency of training and inference could be further improved from several system aspects, including load balancing, communication/computation efficiency, and memory footprint limits. In this work, we present a novel MoESys that boosts efficiency in both large-scale training and inference. Specifically, in the training procedure, the proposed MoESys adopts an Elastic MoE training strategy with 2D prefetch and Fusion communication over Hierarchical storage, so as to enjoy efficient parallelisms. For scalable inference in a single node, especially when the model size is larger than GPU memory, MoESys builds the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference. We carried out extensive experiments to evaluate MoESys, where MoESys successfully trains a Unified Feature Optimization (UFO) model with a Sparsely-Gated Mixture-of-Experts model of 12B parameters in 8 days on 48 A100 GPU cards. The comparison against the state-of-the-art shows that MoESys outperformed DeepSpeed with 33% higher throughput (tokens per second) in training and 13% higher throughput in inference in general. Particularly, under unbalanced MoE Tasks, e.g., UFO, MoESys achieved 64% higher throughput with 18% lower memory footprints.
Pre-trained language models are increasingly important components across multiple information retrieval (IR) paradigms. Late interaction, introduced with the ColBERT model and recently refined in ColBERTv2, is a popular paradigm that holds state-of-the-art status across many benchmarks. To dramatically speed up the search latency of late interaction, we introduce the Performance-optimized Late Interaction Driver (PLAID). Without impacting quality, PLAID swiftly eliminates low-scoring passages using a novel centroid interaction mechanism that treats every passage as a lightweight bag of centroids. PLAID uses centroid interaction as well as centroid pruning, a mechanism for sparsifying the bag of centroids, within a highly-optimized engine to reduce late interaction search latency by up to 7$\times$ on a GPU and 45$\times$ on a CPU against vanilla ColBERTv2, while continuing to deliver state-of-the-art retrieval quality. This allows the PLAID engine with ColBERTv2 to achieve latency of tens of milliseconds on a GPU and tens or just few hundreds of milliseconds on a CPU at large scale, even at the largest scales we evaluate with 140M passages.