Designing a video detection network based on state-of-the-art single-image object detectors may seem like an obvious choice. However, video object detection has extra challenges due to the lower quality of individual frames in a video, and hence the need to include temporal information for high-quality detection results. We design a novel interleaved architecture combining a 2D convolutional network and a 3D temporal network. To explore inter-frame information, we propose feature aggregation based on a temporal network. Our TemporalNet utilizes Appearance-preserving 3D convolution (AP3D) for extracting aligned features in the temporal dimension. Our temporal network functions at multiple scales for better performance, which allows communication between 2D and 3D blocks at each scale and also across scales. Our TemporalNet is a plug-and-play block that can be added to a multi-scale single-image detection network without any adjustments in the network architecture. When TemporalNet is applied to Yolov3 it is real-time with a running time of 35ms/frame on a low-end GPU. Our real-time approach achieves 77.1 % mAP (mean Average Precision) on ImageNet VID 2017 dataset with TemporalNet-4, where TemporalNet-16 achieves 80.9 % mAP which is a competitive result.
We propose the geometry-informed neural operator (GINO), a highly efficient approach to learning the solution operator of large-scale partial differential equations with varying geometries. GINO uses a signed distance function and point-cloud representations of the input shape and neural operators based on graph and Fourier architectures to learn the solution operator. The graph neural operator handles irregular grids and transforms them into and from regular latent grids on which Fourier neural operator can be efficiently applied. GINO is discretization-convergent, meaning the trained model can be applied to arbitrary discretization of the continuous domain and it converges to the continuum operator as the discretization is refined. To empirically validate the performance of our method on large-scale simulation, we generate the industry-standard aerodynamics dataset of 3D vehicle geometries with Reynolds numbers as high as five million. For this large-scale 3D fluid simulation, numerical methods are expensive to compute surface pressure. We successfully trained GINO to predict the pressure on car surfaces using only five hundred data points. The cost-accuracy experiments show a $26,000 \times$ speed-up compared to optimized GPU-based computational fluid dynamics (CFD) simulators on computing the drag coefficient. When tested on new combinations of geometries and boundary conditions (inlet velocities), GINO obtains a one-fourth reduction in error rate compared to deep neural network approaches.
Probabilistic Diffusion Models (PDMs) have recently emerged as a very promising class of generative models, achieving high performance in natural image generation. However, their performance relative to non-natural images, like radar-based satellite data, remains largely unknown. Generating large amounts of synthetic (and especially labelled) satellite data is crucial to implement deep-learning approaches for the processing and analysis of (interferometric) satellite aperture radar data. Here, we leverage PDMs to generate several radar-based satellite image datasets. We show that PDMs succeed in generating images with complex and realistic structures, but that sampling time remains an issue. Indeed, accelerated sampling strategies, which work well on simple image datasets like MNIST, fail on our radar datasets. We provide a simple and versatile open-source https://github.com/thomaskerdreux/PDM_SAR_InSAR_generation to train, sample and evaluate PDMs using any dataset on a single GPU.
Stable diffusion, a generative model used in text-to-image synthesis, frequently encounters resolution-induced composition problems when generating images of varying sizes. This issue primarily stems from the model being trained on pairs of single-scale images and their corresponding text descriptions. Moreover, direct training on images of unlimited sizes is unfeasible, as it would require an immense number of text-image pairs and entail substantial computational expenses. To overcome these challenges, we propose a two-stage pipeline named Any-Size-Diffusion (ASD), designed to efficiently generate well-composed images of any size, while minimizing the need for high-memory GPU resources. Specifically, the initial stage, dubbed Any Ratio Adaptability Diffusion (ARAD), leverages a selected set of images with a restricted range of ratios to optimize the text-conditional diffusion model, thereby improving its ability to adjust composition to accommodate diverse image sizes. To support the creation of images at any desired size, we further introduce a technique called Fast Seamless Tiled Diffusion (FSTD) at the subsequent stage. This method allows for the rapid enlargement of the ASD output to any high-resolution size, avoiding seaming artifacts or memory overloads. Experimental results on the LAION-COCO and MM-CelebA-HQ benchmarks demonstrate that ASD can produce well-structured images of arbitrary sizes, cutting down the inference time by 2x compared to the traditional tiled algorithm.
Large Language Model (LLM) inference consists of two distinct phases - prefill phase which processes the input prompt and decode phase which generates output tokens autoregressively. While the prefill phase effectively saturates GPU compute at small batch sizes, the decode phase results in low compute utilization as it generates one token at a time per request. The varying prefill and decode times also lead to imbalance across micro-batches when using pipeline parallelism, resulting in further inefficiency due to bubbles. We present SARATHI to address these challenges. SARATHI employs chunked-prefills, which splits a prefill request into equal sized chunks, and decode-maximal batching, which constructs a batch using a single prefill chunk and populates the remaining slots with decodes. During inference, the prefill chunk saturates GPU compute, while the decode requests 'piggyback' and cost up to an order of magnitude less compared to a decode-only batch. Chunked-prefills allows constructing multiple decode-maximal batches from a single prefill request, maximizing coverage of decodes that can piggyback. Furthermore, the uniform compute design of these batches ameliorates the imbalance between micro-batches, significantly reducing pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware. For the LLaMA-13B model on A6000 GPU, SARATHI improves decode throughput by up to 10x, and accelerates end-to-end throughput by up to 1.33x. For LLaMa-33B on A100 GPU, we achieve 1.25x higher end-to-end-throughput and up to 4.25x higher decode throughput. When used with pipeline parallelism on GPT-3, SARATHI reduces bubbles by 6.29x, resulting in an end-to-end throughput improvement of 1.91x.
Dynamic computation has emerged as a promising avenue to enhance the inference efficiency of deep networks. It allows selective activation of computational units, leading to a reduction in unnecessary computations for each input sample. However, the actual efficiency of these dynamic models can deviate from theoretical predictions. This mismatch arises from: 1) the lack of a unified approach due to fragmented research; 2) the focus on algorithm design over critical scheduling strategies, especially in CUDA-enabled GPU contexts; and 3) challenges in measuring practical latency, given that most libraries cater to static operations. Addressing these issues, we unveil the Latency-Aware Unified Dynamic Networks (LAUDNet), a framework that integrates three primary dynamic paradigms-spatially adaptive computation, dynamic layer skipping, and dynamic channel skipping. To bridge the theoretical and practical efficiency gap, LAUDNet merges algorithmic design with scheduling optimization, guided by a latency predictor that accurately gauges dynamic operator latency. We've tested LAUDNet across multiple vision tasks, demonstrating its capacity to notably reduce the latency of models like ResNet-101 by over 50% on platforms such as V100, RTX3090, and TX2 GPUs. Notably, LAUDNet stands out in balancing accuracy and efficiency. Code is available at: https://www.github.com/LeapLabTHU/LAUDNet.
As powerful tools for representation learning on graphs, graph neural networks (GNNs) have played an important role in applications including social networks, recommendation systems, and online web services. However, GNNs have been shown to be vulnerable to adversarial attacks, which can significantly degrade their effectiveness. Recent state-of-the-art approaches in adversarial attacks rely on gradient-based meta-learning to selectively perturb a single edge with the highest attack score until they reach the budget constraint. While effective in identifying vulnerable links, these methods are plagued by high computational costs. By leveraging continuous relaxation and parameterization of the graph structure, we propose a novel attack method called Differentiable Graph Attack (DGA) to efficiently generate effective attacks and meanwhile eliminate the need for costly retraining. Compared to the state-of-the-art, DGA achieves nearly equivalent attack performance with 6 times less training time and 11 times smaller GPU memory footprint on different benchmark datasets. Additionally, we provide extensive experimental analyses of the transferability of the DGA among different graph models, as well as its robustness against widely-used defense mechanisms.
Although state-of-the-art (SOTA) SAT solvers based on conflict-driven clause learning (CDCL) have achieved remarkable engineering success, their sequential nature limits the parallelism that may be extracted for acceleration on platforms such as the graphics processing unit (GPU). In this work, we propose FastFourierSAT, a highly parallel hybrid SAT solver based on gradient-driven continuous local search (CLS). This is realized by a novel parallel algorithm inspired by the Fast Fourier Transform (FFT)-based convolution for computing the elementary symmetric polynomials (ESPs), which is the major computational task in previous CLS methods. The complexity of our algorithm matches the best previous result. Furthermore, the substantial parallelism inherent in our algorithm can leverage the GPU for acceleration, demonstrating significant improvement over the previous CLS approaches. We also propose to incorporate the restart heuristics in CLS to improve search efficiency. We compare our approach with the SOTA parallel SAT solvers on several benchmarks. Our results show that FastFourierSAT computes the gradient 100+ times faster than previous prototypes implemented on CPU. Moreover, FastFourierSAT solves most instances and demonstrates promising performance on larger-size instances.
As the size of models and datasets grows, it has become increasingly common to train models in parallel. However, existing distributed stochastic gradient descent (SGD) algorithms suffer from insufficient utilization of computational resources and poor convergence in heterogeneous clusters. In this paper, we propose a delayed synchronous SGD algorithm with adaptive batch size (ABS-SGD) for heterogeneous GPU clusters. In ABS-SGD, workers perform global synchronization to accumulate delayed gradients and use the accumulated delayed gradients to update parameters. While workers are performing global synchronization for delayed gradients, they perform the computation of the next batch without specifying batch size in advance, which lasts until the next global synchronization starts, realizing the full utilization of computational resources. Since the gradient delay is only one iteration, the stale gradient problem can be alleviated. We theoretically prove the convergence of ABS-SGD in heterogeneous clusters. Extensive experiments in three types of heterogeneous clusters demonstrate that ABS-SGD can make full use of computational resources and accelerate model convergence: When training ResNet18 network with 4 workers, ABS-SGD increases the convergence speed by 1.30x on average compared with the best baseline algorithm.
Approximate Nearest Neighbor Search (ANNS) plays a critical role in various disciplines spanning data mining and artificial intelligence, from information retrieval and computer vision to natural language processing and recommender systems. Data volumes have soared in recent years and the computational cost of an exhaustive exact nearest neighbor search is often prohibitive, necessitating the adoption of approximate techniques. The balanced performance and recall of graph-based approaches have more recently garnered significant attention in ANNS algorithms, however, only a few studies have explored harnessing the power of GPUs and multi-core processors despite the widespread use of massively parallel and general-purpose computing. To bridge this gap, we introduce a novel parallel computing hardware-based proximity graph and search algorithm. By leveraging the high-performance capabilities of modern hardware, our approach achieves remarkable efficiency gains. In particular, our method surpasses existing CPU and GPU-based methods in constructing the proximity graph, demonstrating higher throughput in both large- and small-batch searches while maintaining compatible accuracy. In graph construction time, our method, CAGRA, is 2.2~27x faster than HNSW, which is one of the CPU SOTA implementations. In large-batch query throughput in the 90% to 95% recall range, our method is 33~77x faster than HNSW, and is 3.8~8.8x faster than the SOTA implementations for GPU. For a single query, our method is 3.4~53x faster than HNSW at 95% recall.
We present DySample, an ultra-lightweight and effective dynamic upsampler. While impressive performance gains have been witnessed from recent kernel-based dynamic upsamplers such as CARAFE, FADE, and SAPA, they introduce much workload, mostly due to the time-consuming dynamic convolution and the additional sub-network used to generate dynamic kernels. Further, the need for high-res feature guidance of FADE and SAPA somehow limits their application scenarios. To address these concerns, we bypass dynamic convolution and formulate upsampling from the perspective of point sampling, which is more resource-efficient and can be easily implemented with the standard built-in function in PyTorch. We first showcase a naive design, and then demonstrate how to strengthen its upsampling behavior step by step towards our new upsampler, DySample. Compared with former kernel-based dynamic upsamplers, DySample requires no customized CUDA package and has much fewer parameters, FLOPs, GPU memory, and latency. Besides the light-weight characteristics, DySample outperforms other upsamplers across five dense prediction tasks, including semantic segmentation, object detection, instance segmentation, panoptic segmentation, and monocular depth estimation. Code is available at https://github.com/tiny-smart/dysample.
Temporal Interaction Graphs (TIGs) are widely employed to model intricate real-world systems such as financial systems and social networks. To capture the dynamism and interdependencies of nodes, existing TIG embedding models need to process edges sequentially and chronologically. However, this requirement prevents it from being processed in parallel and struggle to accommodate burgeoning data volumes to GPU. Consequently, many large-scale temporal interaction graphs are confined to CPU processing. Furthermore, a generalized GPU scaling and acceleration approach remains unavailable. To facilitate large-scale TIGs' implementation on GPUs for acceleration, we introduce a novel training approach namely Streaming Edge Partitioning and Parallel Acceleration for Temporal Interaction Graph Embedding (SPEED). The SPEED is comprised of a Streaming Edge Partitioning Component (SEP) which addresses space overhead issue by assigning fewer nodes to each GPU, and a Parallel Acceleration Component (PAC) which enables simultaneous training of different sub-graphs, addressing time overhead issue. Our method can achieve a good balance in computing resources, computing time, and downstream task performance. Empirical validation across 7 real-world datasets demonstrates the potential to expedite training speeds by a factor of up to 19.29x. Simultaneously, resource consumption of a single-GPU can be diminished by up to 69%, thus enabling the multiple GPU-based training and acceleration encompassing millions of nodes and billions of edges. Furthermore, our approach also maintains its competitiveness in downstream tasks.
The fashion e-commerce industry has witnessed significant growth in recent years, prompting exploring image-based virtual try-on techniques to incorporate Augmented Reality (AR) experiences into online shopping platforms. However, existing research has primarily overlooked a crucial aspect - the runtime of the underlying machine-learning model. While existing methods prioritize enhancing output quality, they often disregard the execution time, which restricts their applications on a limited range of devices. To address this gap, we propose Distilled Mobile Real-time Virtual Try-On (DM-VTON), a novel virtual try-on framework designed to achieve simplicity and efficiency. Our approach is based on a knowledge distillation scheme that leverages a strong Teacher network as supervision to guide a Student network without relying on human parsing. Notably, we introduce an efficient Mobile Generative Module within the Student network, significantly reducing the runtime while ensuring high-quality output. Additionally, we propose Virtual Try-on-guided Pose for Data Synthesis to address the limited pose variation observed in training images. Experimental results show that the proposed method can achieve 40 frames per second on a single Nvidia Tesla T4 GPU and only take up 37 MB of memory while producing almost the same output quality as other state-of-the-art methods. DM-VTON stands poised to facilitate the advancement of real-time AR applications, in addition to the generation of lifelike attired human figures tailored for diverse specialized training tasks. https://sites.google.com/view/ltnghia/research/DMVTON
Deep learning harnesses massive parallel floating-point processing to train and evaluate large neural networks. Trends indicate that deeper and larger neural networks with an increasing number of parameters achieve higher accuracy than smaller neural networks. This performance improvement, which often requires heavy compute for both training and evaluation, eventually needs to translate well to resource-constrained hardware for practical value. Structured pruning asserts that while large networks enable us to find solutions to complex computer vision problems, a smaller, computationally efficient sub-network can be derived from the large neural network that retains model accuracy but significantly improves computational efficiency. We generalize structured pruning with algorithms for network augmentation, pruning, sub-network collapse and removal. In addition, we demonstrate efficient and stable convergence up to 93% sparsity and 95% FLOPs reduction without loss of inference accuracy using with continuous relaxation matching or exceeding the state of the art for all structured pruning methods. The resulting CNN executes efficiently on GPU hardware without computationally expensive sparse matrix operations. We achieve this with routine automatable operations on classification and segmentation problems using CIFAR-10, ImageNet, and CityScapes datasets with the ResNet and U-NET network architectures.
Fast, reliable shape reconstruction is an essential ingredient in many computer vision applications. Neural Radiance Fields demonstrated that photorealistic novel view synthesis is within reach, but was gated by performance requirements for fast reconstruction of real scenes and objects. Several recent approaches have built on alternative shape representations, in particular, 3D Gaussians. We develop extensions to these renderers, such as integrating differentiable optical flow, exporting watertight meshes and rendering per-ray normals. Additionally, we show how two of the recent methods are interoperable with each other. These reconstructions are quick, robust, and easily performed on GPU or CPU. For code and visual examples, see https://leonidk.github.io/fmb-plus
A code summary is a brief natural language description of source code. Summaries are usually only a single sentence long, and yet form the backbone of developer documentation. A short descriptions such as "changes all visible polygons to the color blue" can give a programmer a high-level idea of what code does without the effort of reading the code itself. Recently, products based on Large Language Models such as ChatGPT have demonstrated a strong ability to write these descriptions automatically. However, to use these tools, programmers must send their code to untrusted third parties for processing (e.g., via an API call). This loss of custody is not acceptable to many organizations. In this paper, we present an alternative: we train an open source model using sample output generated by GPT-3.5 in a process related to knowledge distillation. Our model is small enough (350m parameters) to be run on a single 16gb GPU, yet we show in our evaluation that it is large enough to mimic GPT-3.5 on this task.
We propose the use of conversational GPT models for easy and quick few-shot text classification in the financial domain using the Banking77 dataset. Our approach involves in-context learning with GPT-3.5 and GPT-4, which minimizes the technical expertise required and eliminates the need for expensive GPU computing while yielding quick and accurate results. Additionally, we fine-tune other pre-trained, masked language models with SetFit, a recent contrastive learning technique, to achieve state-of-the-art results both in full-data and few-shot settings. Our findings show that querying GPT-3.5 and GPT-4 can outperform fine-tuned, non-generative models even with fewer examples. However, subscription fees associated with these solutions may be considered costly for small organizations. Lastly, we find that generative models perform better on the given task when shown representative samples selected by a human expert rather than when shown random ones. We conclude that a) our proposed methods offer a practical solution for few-shot tasks in datasets with limited label availability, and b) our state-of-the-art results can inspire future work in the area.
JPEG is one of the most popular image compression methods. It is beneficial to compress those existing JPEG files without introducing additional distortion. In this paper, we propose a deep learning based method to further compress JPEG images losslessly. Specifically, we propose a Multi-Level Parallel Conditional Modeling (ML-PCM) architecture, which enables parallel decoding in different granularities. First, luma and chroma are processed independently to allow parallel coding. Second, we propose pipeline parallel context model (PPCM) and compressed checkerboard context model (CCCM) for the effective conditional modeling and efficient decoding within luma and chroma components. Our method has much lower latency while achieves better compression ratio compared with previous SOTA. After proper software optimization, we can obtain a good throughput of 57 FPS for 1080P images on NVIDIA T4 GPU. Furthermore, combined with quantization, our approach can also act as a lossy JPEG codec which has obvious advantage over SOTA lossy compression methods in high bit rate (bpp$>0.9$).
Despite the recent success of Graph Neural Networks (GNNs), it remains challenging to train GNNs on large-scale graphs due to neighbor explosions. As a remedy, distributed computing becomes a promising solution by leveraging abundant computing resources (e.g., GPU). However, the node dependency of graph data increases the difficulty of achieving high concurrency in distributed GNN training, which suffers from the massive communication overhead. To address it, Historical value approximation is deemed a promising class of distributed training techniques. It utilizes an offline memory to cache historical information (e.g., node embedding) as an affordable approximation of the exact value and achieves high concurrency. However, such benefits come at the cost of involving dated training information, leading to staleness, imprecision, and convergence issues. To overcome these challenges, this paper proposes SAT (Staleness-Alleviated Training), a novel and scalable distributed GNN training framework that reduces the embedding staleness adaptively. The key idea of SAT is to model the GNN's embedding evolution as a temporal graph and build a model upon it to predict future embedding, which effectively alleviates the staleness of the cached historical embedding. We propose an online algorithm to train the embedding predictor and the distributed GNN alternatively and further provide a convergence analysis. Empirically, we demonstrate that SAT can effectively reduce embedding staleness and thus achieve better performance and convergence speed on multiple large-scale graph datasets.
Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements. Although recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization. To tackle this issue, we introduce an Omnidirectionally calibrated Quantization (\textbf{OmniQuant}) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters. OmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights. Operating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization. For instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples. Extensive experiments validate OmniQuant's superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices. Codes are available at \url{https://github.com/OpenGVLab/OmniQuant}.