Reduce Memory Usage Research Articles

High fidelity FEM based on deep learning for arbitrary composite material structure

Due to the outstanding performance, composite materials are widely used and analyzing their properties and designing them based on performance has become a crucial task in the field of many manufacturing industries. Composite materials possess complex multiscale structures, and traditional fine-scale finite element modeling and analysis may lead to severe computational resource challenges. To overcome this difficulty, breakthroughs in key technologies of multiscale accelerated analysis algorithms are required. In this study, an innovative approach based on artificial intelligence and multiscale finite element method is presented. This approach involves partitioning the entire composite material structure into coarse grids that resemble homogenous structures of similar size, providing results consistent to fine-grid finite element analysis. By utilizing CNN for image feature recognition and employing the CGAN adversarial method, coarse-grid equivalent stiffness matrices and multiscale shape functions from completely random microstructures of composite materials can be obtained. Consequently, this enables a rapid response process from microstructure to low-resolution grid to high-resolution physical field, with remarkably accurate physical field results. Moreover, compared to traditional fine-grid finite element methods, this approach significantly reduces memory usage and computation time. This method is applicable to composite materials with varying shaped inclusions, different component properties, and diverse geometric distributions, allowing these materials to perform high-fidelity finite element calculations on coarse grids and predict their mechanical behavior. Furthermore, this breakthrough opens avenues for accelerating the optimization design of composite materials with diverse mechanical functionalities, by employing a bottom-up approach.

Multi-View Separable Residual convolution neural Network for detecting Alzheimer’s disease progression

Alzheimer’s Disease (AD) is a neurodegenerative disorder, the most common form of dementia, characterized by memory loss and cognitive impairments that disrupt daily life. Early detection is crucial for effective treatment. 18F-FDG-PET is the most accurate diagnostic tool, but existing methods often rely on handcrafted or machine learning, risking information loss due to preprocessing. To overcome these limitations, we propose a deep learning approach, leveraging Convolutional Neural Networks (CNNs). We introduce a novel Multi-View Separable Residual CNN (MV-SR-CNN) architecture, capable of processing entire volumes while maintaining spatial complexity similar to 2D CNNs. MV-SR-CNN considers voxel spatial relationships and achieves up to 50 % memory reduction compared to 3D CNNs. We evaluated MV-SR-CNN on a dataset of 540 patients : 191 Control Normal (CN), 145 Early Mild Cognitive Impairment (EMCI), 122 Late Mild Cognitive Impairment (LMCI), and 82 with AD. Additionally, 397 Stable MCI (SMCI) and 61 Progressive MCI (PMCI) cases were included. MV-SR-CNN achieved impressive accuracies of 86.97 % for CN vs. EMCI vs. LMCI vs. AD and 95.73 % for SMCI vs. PMCI classification tasks.

Parsnp 2.0: scalable core-genome alignment for massive microbial datasets.

Since 2016, the number of microbial species with available reference genomes in NCBI has more than tripled. Multiple genome alignment, the process of identifying nucleotides across multiple genomes which share a common ancestor, is used as the input to numerous downstream comparative analysis methods. Parsnp is one of the few multiple genome alignment methods able to scale to the current era of genomic data; however, there has been no major release since its initial release in 2014. To address this gap, we developed Parsnp v2, which significantly improves on its original release. Parsnp v2 provides users with more control over executions of the program, allowing Parsnp to be better tailored for different use-cases. We introduce a partitioning option to Parsnp, which allows the input to be broken up into multiple parallel alignment processes which are then combined into a final alignment. The partitioning option can reduce memory usage by over 4× and reduce runtime by over 2×, all while maintaining a precise core-genome alignment. The partitioning workflow is also less susceptible to complications caused by assembly artifacts and minor variation, as alignment anchors only need to be conserved within their partition and not across the entire input set. We highlight the performance on datasets involving thousands of bacterial and viral genomes. Parsnp v2 is available at https://github.com/marbl/parsnp.

SplitDF: Splitting Dataframes for Memory-Efficient Data Analysis

Dataframe is a popular construct in data analysis libraries that offers a tabular view of the data. However, data within a dataframe often has redundancy, which can lead to high memory utilization of data analysis libraries. Inspired by the process of normalization in relational database systems, we propose a technique called splitting that can be applied to tabular data to reduce redundancy. Splitting involves performing lossless join decomposition by explicitly adding joining keys, and unlike normalization, splitting can be applied to tabular data without the need to perform functional dependency discovery. A split dataframe provides the same unified tabular view to the data, while internally operating on split data to improve memory efficiency. We develop SplitDF, an implementation of split dataframes in Ibis for DuckDB backend, which enables data analysis on split data with minimal changes to the Ibis API. Generation of split tabular data is automated using an algorithm SplitGen implemented in Velox. In our analysis involving ten handwritten notebooks running on SplitDF, we observe a reduction in memory usage of 19--61% when operating on split data as compared to operating on original data.

Indoor Navigation in Facilities with Repetitive Structures.

Most facilities are structured in a repetitive manner. In this paper, we propose an algorithm and its partial implementation for a cellular guide in such facilities without GPS use. The complete system is based on iBeacons-like components, which operate on BLE technology, and their integration into a navigation application. We assume that the user's location is determined with sufficient accuracy. Our main goal revolves around leveraging the repetitive structure of the given facility to optimize navigation in terms of storage requirements, energy efficiency in the cellular device, algorithmic complexity, and other aspects. To the best of our knowledge, there is no prior experience in addressing this specific aim. In order to provide high performance in real time, we rely on optimal saving and the use of pre-calculated and stored navigation sub-routes. Our implementation seamlessly integrates iBeacon communications, a pre-defined indoor map, diverse data structures for efficient information storage, and a user interface, all working cohesively under a single supervision. Each module can be considered, developed, and improved independently. The approach is mainly directed to places, such as passenger ships, hotels, colleges, and so on. Because of the fact that there are "replicated" parts on different floors, stored once and used for multiple routes, we reduce the amount of information that must be stored, thus helping to reduce memory usage and as a result, yielding a better running time and energy consumption.

Flow field and particle dispersion in a naturally ventilated model building: Application of large eddy simulation based on parallel lattice Boltzmann method

Flow field analysis and pollutant particles tracking in a natural ventilation flow are computationally intensive due to the coupled outdoor and indoor airflow, as well as the existence of different turbulence scales, and large numbers of particles. As a breakthrough approach, an explicit and parallel algorithm based on lattice Boltzmann method (LBM) solver is developed to simulate transient particulate flow indoor and outdoor of a model building. Since flow analysis in natural ventilation requires high-order CFD methods, large eddy simulation (LES) by the wall-adapting local eddy-viscosity (WALE) sub-grid scales model is employed. Also, one-step collision for reduction of memory usage, domain decomposition method for parallel processing of both flow and particles, and Lagrangian tracking for following particles are implemented into the developed algorithm. Investigation of grid independency shows that medium lattice resolution provides an average mean deviation equal to 18% and 34% for horizontal and vertical centerline of model building, respectively. In addition, when compared to finite volume method (FVM), LBM requires 2.7 times higher lattice resolution to achieve acceptable accuracy. However, it approximately speeds simulation duration up to 3 times. Moreover, 100 seconds of parallel particle tracking lasts almost 800 minutes on a simple desktop.

Empowering the SDM-RDFizer tool for scaling up to complex knowledge graph creation pipelines1

The significant increase in data volume in recent years has prompted the adoption of knowledge graphs as valuable data structures for integrating diverse data and metadata. However, this surge in data availability has brought to light challenges related to standardization, interoperability, and data quality. Knowledge graph creation faces complexities from large data volumes, data heterogeneity, and high duplicate rates. This work addresses these challenges and proposes data management techniques to scale up the creation of knowledge graphs specified using the RDF Mapping Language (RML). These techniques are integrated into SDM-RDFizer, transforming it into a two-fold solution designed to address the complexities of generating knowledge graphs. Firstly, we introduce a reordering approach for RML triples maps, prioritizing the evaluation of the most selective maps first to reduce memory usage. Secondly, we employ an RDF compression strategy, along with optimized data structures and novel operators, to prevent the generation of duplicate RDF triples and optimize the execution of RML operators. We assess the performance of SDM-RDFizer through established benchmarks. The evaluation showcases the effectiveness of SDM-RDFizer compared to state-of-the-art RML engines, emphasizing the benefits of our techniques. Furthermore, the paper presents real-world projects where SDM-RDFizer has been utilized, providing insights into the advantages of declaratively defining knowledge graphs and efficiently executing these specifications using this engine.

StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting

Stock price forecasting is a fundamental yet challenging task in quantitative investment. Various researchers have developed a combination of neural network models (e.g., RNNs, GNNs, Transformers) for capturing complex indicator, temporal and stock correlations of the stock data.While complex architectures are highly expressive, they are often difficult to optimize and the performances are often compromised by the limited stock data. In this paper, we propose a simple MLP-based architecture named StockMixer which is easy to optimize and enjoys strong predictive performance. StockMixer performs indicator mixing, followed by time mixing, and finally stock mixing. Unlike the standard MLP-based mixing, we devise the time mixing to exchange multi-scale time patch information and realize the stock mixing by exploiting stock-to-market and market-to-stock influences explicitly. Extensive experiments on real stock benchmarks demonstrate our proposed StockMixer outperforms various state-of-the-art forecasting methods with a notable margin while reducing memory usage and runtime cost.Code is available at https://github.com/SJTU-Quant/StockMixer.

Darwin3: a large-scale neuromorphic chip with a novel ISA and on-chip learning.

Spiking neural networks (SNNs) are gaining increasing attention for their biological plausibility and potential for improved computational efficiency. To match the high spatial-temporal dynamics in SNNs, neuromorphic chips are highly desired to execute SNNs in hardware-based neuron and synapse circuits directly. This paper presents a large-scale neuromorphic chip named Darwin3 with a novel instruction set architecture, which comprises 10 primary instructions and a few extended instructions. It supports flexible neuron model programming and local learning rule designs. The Darwin3 chip architecture is designed in a mesh of computing nodes with an innovative routing algorithm. We used a compression mechanism to represent synaptic connections, significantly reducing memory usage. The Darwin3 chip supports up to 2.35 million neurons, making it the largest of its kind on the neuron scale. The experimental results showed that the code density was improved by up to 28.3× in Darwin3, and that the neuron core fan-in and fan-out were improved by up to 4096× and 3072× by connection compression compared to the physical memory depth. Our Darwin3 chip also provided memory saving between 6.8× and 200.8× when mapping convolutional spiking neural networks onto the chip, demonstrating state-of-the-art performance in accuracy and latency compared to other neuromorphic chips.

A Scheduler for Smart Home Appliances Based on a Novel Concept of Tariff Space.

The background of this work is related to the scheduling of household appliances, taking into account variations in energy costs during the day from official Brazilian domestic tariffs: constant and white. The white tariff can reach an average price of around 17% lower than the constant, but charges twice its value at peak hours. In addition to cost reduction, we propose a methodology to reduce user discomfort due to time-shifting of controllable devices, presenting a balanced solution through the analytical analysis of a new method referred to as tariff space, derived from white tariff posts. To achieve this goal, we explore the geometric properties of the movement of devices through the tariff space (geometric locus of the load), over which we can define a limited region in which the cost of a load under the white tariff will be equal to or less than the constant tariff. As a trial for the efficiency of this new methodology, we collected some benchmarks (such as execution time and memory usage) against a classic multi-objective algorithm (hierarchical) available in the language portfolio in which the project has been executed (the Julia language). As a result, while both methodologies yield similar results, the approach presented in this article demonstrates a significant reduction in processing time and memory usage, which could lead to the future implementation of the solution in a simple, low-cost embedded system like an ARM cortex M.

A novel Swin transformer approach utilizing residual multi-layer perceptron for diagnosing brain tumors in MRI images

Serious consequences due to brain tumors necessitate a timely and accurate diagnosis. However, obstacles such as suboptimal imaging quality, issues with data integrity, varying tumor types and stages, and potential errors in interpretation hinder the achievement of precise and prompt diagnoses. The rapid identification of brain tumors plays a pivotal role in ensuring patient safety. Deep learning-based systems hold promise in aiding radiologists to make diagnoses swiftly and accurately. In this study, we present an advanced deep learning approach based on the Swin Transformer. The proposed method introduces a novel Hybrid Shifted Windows Multi-Head Self-Attention module (HSW-MSA) along with a rescaled model. This enhancement aims to improve classification accuracy, reduce memory usage, and simplify training complexity. The Residual-based MLP (ResMLP) replaces the traditional MLP in the Swin Transformer, thereby improving accuracy, training speed, and parameter efficiency. We evaluate the Proposed-Swin model on a publicly available brain MRI dataset with four classes, using only test data. Model performance is enhanced through the application of transfer learning and data augmentation techniques for efficient and robust training. The Proposed-Swin model achieves a remarkable accuracy of 99.92%, surpassing previous research and deep learning models. This underscores the effectiveness of the Swin Transformer with HSW-MSA and ResMLP improvements in brain tumor diagnosis. This method introduces an innovative diagnostic approach using HSW-MSA and ResMLP in the Swin Transformer, offering potential support to radiologists in timely and accurate brain tumor diagnosis, ultimately improving patient outcomes and reducing risks.

Optimal Re-Materialization Strategies for Heterogeneous Chains: How to Train Deep Neural Networks with Limited Memory

Training in Feed Forward Deep Neural Networks is a memory-intensive operation which is usually performed on GPUs with limited memory capacities. This may force data scientists to limit the depth of the models or the resolution of the input data if data does not fit in the GPU memory. The re-materialization technique, whose idea comes from the checkpointing strategies developed in the Automatic Differentiation literature, allows data scientists to limit the memory requirements related to the storage of intermediate data (activations), at the cost of an increase in the computational cost. This paper introduces a new strategy of re-materialization of activations that significantly reduces memory usage. It consists in selecting which activations are saved and which activations are deleted during the forward phase, and then recomputing the deleted activations when they are needed during the backward phase. We propose an original computation model that combines two types of activation savings: either only storing the layer inputs, or recording the complete history of operations that produced the outputs. This paper focuses on the fully heterogeneous case, where the computation time and the memory requirement of each layer is different. We prove that finding the optimal solution is NP-hard and that classical techniques from Automatic Differentiation literature do not apply. Moreover, the classical assumption of memory persistence of materialized activations, used to simplify the search of optimal solutions, does not hold anymore. Thus, we propose a weak memory persistence property and provide a Dynamic Program to compute the optimal sequence of computations. This algorithm is made available through the Rotor software, a PyTorch plug-in dealing with any network consisting of a sequence of layers, each of them having an arbitrarily complex structure. Through extensive experiments, we show that our implementation consistently outperforms existing re-materialization approaches for a large class of networks, image sizes and batch sizes.

PMAlloc: A Holistic Approach to Improving Persistent Memory Allocation

Persistent memory allocation is a fundamental building block for developing high-performance and in-memory applications. Existing persistent memory allocators suffer from many performance issues. First, they may introduce repeated cache line flushes and small random accesses in persistent memory for their poor heap metadata management. Second, they use static slab segregation resulting in a dramatic increase in memory consumption when allocation request size is changed. Third, they are not aware of NUMA effect, leading to remote persistent memory accesses in memory allocation and deallocation processes. In this paper, we design a novel allocator, named PMAlloc, to solve the above issues simultaneously. (1) PMAlloc eliminates cache line reflushes by mapping contiguous data blocks in slabs to interleaved metadata entries stored in different cache lines. (2) It writes small metadata units to a persistent bookkeeping log in a sequential pattern to remove random heap metadata accesses in persistent memory. (3) Instead of using static slab segregation, it supports slab morphing, which allows slabs to be transformed between size classes to significantly improve slab usage. (4) It uses a local-first allocation policy to avoid allocating remote memory blocks. And it supports a two-phase deallocation mechanism including recording and synchronization to minimize the number of remote memory access in the deallocation. PMAlloc is complementary to the existing consistency models. Results on 6 benchmarks demonstrate that PMAlloc improves the performance of state-of-the-art persistent memory allocators by up to 6.4x and 57x for small and large allocations, respectively. PMAlloc with NUMA optimizations brings a 2.9x speedup in multi-socket evaluation and is up to 36x faster than other persistent memory allocators. Using PMAlloc reduces memory usage by up to 57.8%. Besides, we integrate PMAlloc in a persistent FPTree. Compared to the state-of-the-art allocators, PMAlloc improves the performance of this application by up to 3.1x.

2-D seismic wave propagation using the distributional finite-difference method: further developments and potential for global seismology

SUMMARY We present a time-domain distributional finite-difference scheme based on the Lebedev staggered grid for the numerical simulation of wave propagation in acoustic and elastic media. The central aspect of the proposed method is the representation of the stresses and displacements with different sets of B-splines functions organized according to the staggered grid. The distributional finite-difference approach allows domain-decomposition, heterogeneity of the medium, curvilinear mesh, anisotropy, non-conformal interfaces, discontinuous grid and fluid–solid interfaces. Numerical examples show that the proposed scheme is suitable to model wave propagation through the Earth, where sharp interfaces separate large, relatively homogeneous layers. A few domains or elements are sufficient to represent the Earth’s internal structure without relying on advanced meshing techniques. We compare seismograms obtained with the proposed scheme and the spectral element method, and we show that our approach offers superior accuracy, reduced memory usage, and comparable efficiency.

TeRa: Ternary and Range based packet classification engine

This work proposes a novel approach to the hardware implementation of packet classification in ASICs, using NAND-NOR logic at each stage. The proposed design utilizes modified ternary encoding to process the prefix field, resulting in a two-level NAND-NOR logic for prefix processing. The field representation based on classbench rules helps reduce memory usage by almost 45% compared to conventional prefix representations. Additionally, an efficient range matching solution is implemented using a carry tree logic that relies on 1’s and 2’s Complement subtraction. The integration of a carry-based range comparator enhances hardware optimization for range processing without the need for prefix conversion. Additionally, match inversion logic streamlines the processing of exceptional or inverse fields without incurring extra hardware overhead. This work also presents a specialized logic circuit for ternary and range matching, complemented by a specialized priority grouping technique. The proposed architecture, including TYPE1 and TYPE2, achieves throughput rates of 9.9 BPPS and 6.6 BPPS, respectively, while supporting both best-match and multi-match addresses within the same hardware environment.

Accelerated generation of holograms with ultra-low memory symmetrically high-compressed look-up table

Computer-generated holography technology has been widely applied, and as research in this field deepens, the demand for memory and computational power in small AR and VR devices continues to increase. This paper presents a hologram generation method, i.e., a symmetrically high-compressed look-up table method, which can reduce memory usage by 50%. In offline computing, half of the basic horizontal and vertical modulation factors are stored, halving the memory requirements without affecting inline speed. Currently, its potential extends to various holographic applications, including the production of optical diffraction elements.

FastSegFormer: A knowledge distillation-based method for real-time semantic segmentation of surface defects in navel oranges

Navel oranges are valued citrus fruits with a strong market presence, and detecting defects is crucial in their sorting due to common diseases and abnormalities during growth and transport. Deep learning, particularly semantic segmentation, is revolutionizing the fruit sorting industry by overcoming the limitations of traditional defect detection and enhancing the accuracy of classifying complex defects in navel oranges. The FastSegFormer network, enabling real-time fruit defect detection, addresses this challenge with our introduced Multi-scale Pyramid (MSP) module for its architecture and a semi-resolution reconstruction branch post-feature fusion. We suggested a multi-resolution knowledge distillation strategy to further increase the network’s segmentation accuracy. We developed a navel orange defect segmentation dataset, trained, and evaluated our FastSegFormer-E model, designed for memory-constrained devices. It outperforms ENet by 3.15%, achieving a mIoU of 88.78% on the test set. The FastSegFormer-P model, tailored for high-speed detection, was tested on the mid-range RTX3060 graphics card, surpassing ENet by 3.7% with a mIoU of 89.33% and reaching 108 frames/s. The results demonstrate that the FastSegFormer-E model attains enhanced detection accuracy with reduced memory usage, whereas the FastSegFormer-P model stands out by striking an optimal balance between top-tier detection accuracy and rapid processing speed. Deploying the algorithm system on the same platform as pipeline sorting, 20 frame/s was achieved on a Jetson Nano with very low computational power. The model significantly improves the detection of subtle and intricate edge defects, achieving real-time speeds. Our proposed algorithm enhances the fineness of fruit sorting, resolves the limitation of existing algorithms that apply to a narrow range of fruit sorting scenarios, and provides an efficient and accurate solution for large-scale navel orange defect detection.

In situ adaptive tabulation of vapor-liquid equilibrium solutions for multi-component high-pressure transcritical flows with phase change

Vapor-liquid equilibrium (VLE) is a family of first-principled thermodynamic models for transcritical multiphase flows, which can accurately capture the phase transitions at high-pressure conditions that are difficult to deal with using other models. However, VLE-based computational fluid dynamics (CFD) simulation is computationally very expensive for multi-component systems, which severely limits its applications to real-world systems. In this work, we developed a new ISAT-VLE method based on the in situ adaptive tabulation (ISAT) method to improve the computational efficiency of VLE-based CFD simulation with reduced memory usage. We developed several ISAT-VLE solvers for both fully conservative (FC) and double flux (DF) schemes. New methods are proposed to delete redundant records in the ISAT-VLE table and the ISAT-VLE method performance is further improved. To improve the convergence of the VLE solvers, a modified initial guess for equilibrium constant is also introduced. Simulations of high-pressure transcritical two-phase temporal mixing layers and shock-droplet interaction were conducted using the ISAT-VLE CFD solvers. The simulation results show that the new method obtains a speed-up factor approximately from 10 to 60 and the ISAT errors can be controlled within 1%. The shock-droplet interaction results show that the DF scheme can achieve a higher speed-up factor than the FC scheme. The two sets of simulations exhibit the phase separation at high-pressure conditions. It was found that even at supercritical pressures with respect to each component, the droplet surface could still be in a subcritical two-phase state, because the mixture critical pressure is often significantly higher than each component and hence triggers phase separation. In addition, a shock wave could partially or completely convert the droplet surface from a subcritical two-phase state to a single-phase state by raising temperature and pressure.

Ship Infrared Automatic Target Recognition Based on Bipartite Graph Recommendation: A Model-Matching Method

Deep learning technology has greatly propelled the development of intelligent and information-driven research on ship infrared automatic target recognition (SIATR). In future scenarios, there will be various recognition models with different mechanisms to choose from. However, in complex and dynamic environments, ship infrared (IR) data exhibit rich feature space distribution, resulting in performance variations among SIATR models, thus preventing the existence of a universally superior model for all recognition scenarios. In light of this, this study proposes a model-matching method for SIATR tasks based on bipartite graph theory. This method establishes evaluation criteria based on recognition accuracy and feature learning credibility, uncovering the underlying connections between IR attributes of ships and candidate models. The objective is to selectively recommend the optimal candidate model for a given sample, enhancing the overall recognition performance and applicability of the model. We separately conducted tests for the optimization of accuracy and credibility on high-fidelity simulation data, achieving Accuracy and EDMS (our credibility metric) of 95.86% and 0.7781. Our method improves by 1.06% and 0.0274 for each metric compared to the best candidate models (six in total). Subsequently, we created a recommendation system that balances two tasks, resulting in improvements of 0.43% (accuracy) and 0.0071 (EDMS). Additionally, considering the relationship between model resources and performance, we achieved a 28.35% reduction in memory usage while realizing enhancements of 0.33% (accuracy) and 0.0045 (EDMS).

MCMC: Multi-Constrained Model Compression via One-Stage Envelope Reinforcement Learning.

Model compression methods are being developed to bridge the gap between the massive scale of neural networks and the limited hardware resources on edge devices. Since most real-world applications deployed on resource-limited hardware platforms typically have multiple hardware constraints simultaneously, most existing model compression approaches that only consider optimizing one single hardware objective are ineffective. In this article, we propose an automated pruning method called multi-constrained model compression (MCMC) that allows for the optimization of multiple hardware targets, such as latency, floating point operations (FLOPs), and memory usage, while minimizing the impact on accuracy. Specifically, we propose an improved multi-objective reinforcement learning (MORL) algorithm, the one-stage envelope deep deterministic policy gradient (DDPG) algorithm, to determine the pruning strategy for neural networks. Our improved one-stage envelope DDPG algorithm reduces exploration time and offers greater flexibility in adjusting target priorities, enhancing its suitability for pruning tasks. For instance, on the visual geometry group (VGG)-16 network, our method achieved an 80% reduction in FLOPs, a 2.31× reduction in memory usage, and a 1.92× acceleration, with an accuracy improvement of 0.09% compared with the baseline. For larger datasets, such as ImageNet, we reduced FLOPs by 50% for MobileNet-V1, resulting in a 4.7× faster speed and 1.48× memory compression, while maintaining the same accuracy. When applied to edge devices, such as JETSON XAVIER NX, our method resulted in a 71% reduction in FLOPs for MobileNet-V1, leading to a 1.63× faster speed, 1.64× memory compression, and an accuracy improvement.