Approximating radiance fields with volumetric grids is one of promising directions for improving NeRF, represented by methods like Plenoxels and DVGO, which achieve super-fast training convergence and real-time rendering. However, these methods typically require a tremendous storage overhead, costing up to hundreds of megabytes of disk space and runtime memory for a single scene. We address this issue in this paper by introducing a simple yet effective framework, called vector quantized radiance fields (VQRF), for compressing these volume-grid-based radiance fields. We first present a robust and adaptive metric for estimating redundancy in grid models and performing voxel pruning by better exploring intermediate outputs of volumetric rendering. A trainable vector quantization is further proposed to improve the compactness of grid models. In combination with an efficient joint tuning strategy and post-processing, our method can achieve a compression ratio of 100$\times$ by reducing the overall model size to 1 MB with negligible loss on visual quality. Extensive experiments demonstrate that the proposed framework is capable of achieving unrivaled performance and well generalization across multiple methods with distinct volumetric structures, facilitating the wide use of volumetric radiance fields methods in real-world applications. Code Available at \url{//github.com/AlgoHunt/VQRF}
相關內容
We propose a matrix-free solver for the numerical solution of the cardiac electrophysiology model consisting of the monodomain nonlinear reaction-diffusion equation coupled with a system of ordinary differential equations for the ionic species. Our numerical approximation is based on the high-order Spectral Element Method (SEM) to achieve accurate numerical discretization while employing a much smaller number of Degrees of Freedom than first-order Finite Elements. We combine vectorization with sum-factorization, thus allowing for a very efficient use of high-order polynomials in a high performance computing framework. We validate the effectiveness of our matrix-free solver in a variety of applications and perform different electrophysiological simulations ranging from a simple slab of cardiac tissue to a realistic four-chamber heart geometry. We compare SEM to SEM with Numerical Integration (SEM-NI), showing that they provide comparable results in terms of accuracy and efficiency. In both cases, increasing the local polynomial degree $p$ leads to better numerical results and smaller computational times than reducing the mesh size $h$. We also implement a matrix-free Geometric Multigrid preconditioner that results in a comparable number of linear solver iterations with respect to a state-of-the-art matrix-based Algebraic Multigrid preconditioner. As a matter of fact, the matrix-free solver proposed here yields up to 45$\times$ speed-up with respect to a conventional matrix-based solver.
Psychoacoustic experiments have shown that directional properties of the direct sound, salient reflections, and the late reverberation of an acoustic room response can have a distinct influence on the auditory perception of a given room. Spatial room impulse responses (SRIRs) capture those properties and thus are used for direction-dependent room acoustic analysis and virtual acoustic rendering. This work proposes a subspace method that decomposes SRIRs into a direct part, which comprises the direct sound and the salient reflections, and a residual, to facilitate enhanced analysis and rendering methods by providing individual access to these components. The proposed method is based on the generalized singular value decomposition and interprets the residual as noise that is to be separated from the other components of the reverberation. Large generalized singular values are attributed to the direct part, which is then obtained as a low-rank approximation of the SRIR. By advancing from the end of the SRIR toward the beginning while iteratively updating the residual estimate, the method adapts to spatio-temporal variations of the residual. The method is evaluated using a spatio-spectral error measure and simulated SRIRs of different rooms, microphone arrays, and ratios of direct sound to residual energy. The proposed method creates lower errors than existing approaches in all tested scenarios, including a scenario with two simultaneous reflections. A case study with measured SRIRs shows the applicability of the method under real-world acoustic conditions. A reference implementation is provided.
This work focuses on reducing neural network size, which is a major driver of neural network execution time, power consumption, bandwidth, and memory footprint. A key challenge is to reduce size in a manner that can be exploit-ed readily for efficient training and inference without the need for specialized hardware. We propose Self-Compression: a simple, general method that simultaneously achieves two goals: (1) removing redundant weights, and (2) reducing the number of bits required to represent the remaining weights. This is achieved using a generalized loss function to minimize overall network size. In our ex-periments we demonstrate floating point accuracy with as few as 3% of the bits and 18% of the weights remaining in the network.
JPEG image compression algorithm is a widely used technique for image size reduction in edge and cloud computing settings. However, applying such lossy compression on images processed by deep neural networks can lead to significant accuracy degradation. Inspired by the curriculum learning paradigm, we propose a training approach called curriculum pre-training (CPT) for crowd counting on compressed images, which alleviates the drop in accuracy resulting from lossy compression. We verify the effectiveness of our approach by extensive experiments on three crowd counting datasets, two crowd counting DNN models and various levels of compression. The proposed training method is not overly sensitive to hyper-parameters, and reduces the error, particularly for heavily compressed images, by up to 19.70%.
Agent-Based Models are very useful for simulation of physical or social processes, such as the spreading of a pandemic in a city. Such models proceed by specifying the behavior of individuals (agents) and their interactions, and parameterizing the process of infection based on such interactions based on the geography and demography of the city. However, such models are computationally very expensive, and the complexity is often linear in the total number of agents. This seriously limits the usage of such models for simulations, which often have to be run hundreds of times for policy planning and even model parameter estimation. An alternative is to develop an emulator, a surrogate model that can predict the Agent-Based Simulator's output based on its initial conditions and parameters. In this paper, we discuss a Deep Learning model based on Dilated Convolutional Neural Network that can emulate such an agent based model with high accuracy. We show that use of this model instead of the original Agent-Based Model provides us major gains in the speed of simulations, allowing much quicker calibration to observations, and more extensive scenario analysis. The models we consider are spatially explicit, as the locations of the infected individuals are simulated instead of the gross counts. Another aspect of our emulation framework is its divide-and-conquer approach that divides the city into several small overlapping blocks and carries out the emulation in them parallelly, after which these results are merged together. This ensures that the same emulator can work for a city of any size, and also provides significant improvement of time complexity of the emulator, compared to the original simulator.
We propose a Gaussian variational inference framework for the motion planning problem. In this framework, motion planning is formulated as an optimization over the distribution of the trajectories to approximate the desired trajectory distribution by a tractable Gaussian distribution. Equivalently, the proposed framework can be viewed as a standard motion planning with an entropy regularization. Thus, the solution obtained is a transition from an optimal deterministic solution to a stochastic one, and the proposed framework can recover the deterministic solution by controlling the level of stochasticity. To solve this optimization, we adopt the natural gradient descent scheme. The sparsity structure of the proposed formulation induced by factorized objective functions is further leveraged to improve the scalability of the algorithm. We evaluate our method on several robot systems in simulated environments, and show that it achieves collision avoidance with smooth trajectories, and meanwhile brings robustness to the deterministic baseline results, especially in challenging environments and tasks.
Emerging smart grid applications analyze large amounts of data collected from millions of meters and systems to facilitate distributed monitoring and real-time control tasks. However, current parallel data processing systems are designed for common applications, unaware of the massive volume of the collected data, causing long data transfer delay during the computation and slow response time of smart grid systems. A promising direction to reduce delay is to jointly schedule computation tasks and data transfers. We identify that the smart grid data analytic jobs require the intermediate data among different computation stages to be transmitted orderly to avoid network congestion. This new feature prevents current scheduling algorithms from being efficient. In this work, an integrated computing and communication task scheduling scheme is proposed. The mathematical formulation of smart grid data analytic jobs scheduling problem is given, which is unsolvable by existing optimization methods due to the strongly coupled constraints. Several techniques are combined to linearize it for adapting the Branch and Cut method. Based on the topological information in the job graph, the Topology Aware Branch and Cut method is further proposed to speed up searching for optimal solutions. Numerical results demonstrate the effectiveness of the proposed method.
As soon as abstract mathematical computations were adapted to computation on digital computers, the problem of efficient representation, manipulation, and communication of the numerical values in those computations arose. Strongly related to the problem of numerical representation is the problem of quantization: in what manner should a set of continuous real-valued numbers be distributed over a fixed discrete set of numbers to minimize the number of bits required and also to maximize the accuracy of the attendant computations? This perennial problem of quantization is particularly relevant whenever memory and/or computational resources are severely restricted, and it has come to the forefront in recent years due to the remarkable performance of Neural Network models in computer vision, natural language processing, and related areas. Moving from floating-point representations to low-precision fixed integer values represented in four bits or less holds the potential to reduce the memory footprint and latency by a factor of 16x; and, in fact, reductions of 4x to 8x are often realized in practice in these applications. Thus, it is not surprising that quantization has emerged recently as an important and very active sub-area of research in the efficient implementation of computations associated with Neural Networks. In this article, we survey approaches to the problem of quantizing the numerical values in deep Neural Network computations, covering the advantages/disadvantages of current methods. With this survey and its organization, we hope to have presented a useful snapshot of the current research in quantization for Neural Networks and to have given an intelligent organization to ease the evaluation of future research in this area.
Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.
Deep neural networks have achieved remarkable success in computer vision tasks. Existing neural networks mainly operate in the spatial domain with fixed input sizes. For practical applications, images are usually large and have to be downsampled to the predetermined input size of neural networks. Even though the downsampling operations reduce computation and the required communication bandwidth, it removes both redundant and salient information obliviously, which results in accuracy degradation. Inspired by digital signal processing theories, we analyze the spectral bias from the frequency perspective and propose a learning-based frequency selection method to identify the trivial frequency components which can be removed without accuracy loss. The proposed method of learning in the frequency domain leverages identical structures of the well-known neural networks, such as ResNet-50, MobileNetV2, and Mask R-CNN, while accepting the frequency-domain information as the input. Experiment results show that learning in the frequency domain with static channel selection can achieve higher accuracy than the conventional spatial downsampling approach and meanwhile further reduce the input data size. Specifically for ImageNet classification with the same input size, the proposed method achieves 1.41% and 0.66% top-1 accuracy improvements on ResNet-50 and MobileNetV2, respectively. Even with half input size, the proposed method still improves the top-1 accuracy on ResNet-50 by 1%. In addition, we observe a 0.8% average precision improvement on Mask R-CNN for instance segmentation on the COCO dataset.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191