Dynamic sparsity and model feature learning enhanced training for convolutional neural network-pruning

Ruan, Xinbo; Hu, Weiming; Liu, Yufan; Li, Bing

doi:10.1360/sst-2021-0088

Cited by 6 publications

(5 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As the last layer of one-dimensional convolutional neural network [17], the SoftMax layer is mainly used to output the probability of trade volume of different Sino-Korean cultural products and predict the trade situation of Sino-Korean cultural products, as shown in Equation ( 10):…”

Section: Construction Of One-dimensional Convolutional Neural Networkmentioning

confidence: 99%

“…(2) Due to the influence of various factors and the interaction between them, the changes in trade in cultural products are relatively random, resulting in significant differences between the historical data of trade in cultural products. Therefore, the historical data of trade in cultural products are standardized according to Equation (17) to keep the trade in cultural products within the range of   0,1 . (3) Some data are selected from the preprocessed historical data of Sino-Korean cultural product trade to form a training sample, and the remaining historical data of Sino-Korean cultural product trade amount to form a test sample.…”

Section: Prediction Processmentioning

confidence: 99%

See 1 more Smart Citation

Big data integration technology based Prediction of Sino-Korean cultural products trade

Liu

2023

Preprint

View full text Add to dashboard Cite

In order to better manage the trade in cultural products between China and South Korea, and solve the trade deficit problem, this paper focuses on the prediction method of trade in cultural products between China and South Korea based on big data integration technology. Various types of data in the process of trade in cultural products between China and South Korea are collected, to estimate and classify the controllable correlation index big data that affects trade prediction online, and determine the direction of trade data mining; Design association rule functions to determine available data sources. The available data are clustered using a hybrid data clustering algorithm based on integration technology and spectral clustering technology; Data features of Sino-Korean cultural product trade are mined through data matching principles, semantic analysis, and other methods. Taking trade data features as learning samples for trade forecasting, big data integration techniques are used, namely one-dimensional convolutional neural networks and support vector machines, respectively, to model and predict the trade in cultural products between China and South Korea, so as to obtain the final prediction results of trade in cultural products between China and South Korea through reasonable weighting. The experimental results show that under the conditions of setting the parameters of the prediction model, the method can accurately predict the trade situation of cultural products between China and South Korea. During the prediction process, the credibility measurement value and controllable correlation degree are always higher than 19 and 12.5, and the uncertainty discrimination degree and error coefficient are always lower than 12 and 6.

show abstract

Section: Construction Of One-dimensional Convolutional Neural Networkmentioning

confidence: 99%

Section: Prediction Processmentioning

confidence: 99%

Big data integration technology based Prediction of Sino-Korean cultural products trade

Liu

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…The calculation and data flow in FPGA can be uniformly allocated through various instruction segments. The length L of each segment can be macro-defined according according to (13). N i represents the number of species included in the i-th segment.…”

Section: Folding Designmentioning

confidence: 99%

“…Additionally, quantization [13] can be used to reduce model sizes and hardware resource consumption, such as replacing original 32-bit floating-point operations with lower precision fixed-point numbers like 8-bit or 16-bit. Jacob et al [15] proposed an integer quantization method, which uniformly quantizes both weight and activation to 8-bit.…”

Section: Introductionmentioning

confidence: 99%

Research on Efficient CNN Acceleration Through Mixed Precision Quantization: A Comprehensive Methodology

He,

Liu,

Tahir

et al. 2023

IJACSA

View full text Add to dashboard Cite

To overcome challenges associated with deploying Convolutional Neural Networks (CNNs) on edge computing devices with limited memory and computing resources, we propose a mixed-precision CNN calculation method on a Field Programmable Gate Array (FPGA). This approach involves a collaborative design encompassing both software and hardware aspects. Initially, we devised a CNN quantization method tailored for the fixed-point operation characteristics of FPGA, addressing the computational challenges posed by floating-point parameters. We introduce a bit-width strategy search algorithm that assigns bit-widths to each layer based on CNN loss variation induced by quantization. Through retraining, this strategy mitigates the degradation in CNN inference accuracy. For FPGA acceleration design, we employ a flow processing architecture with multiple Processing Elements (PEs) to support mixed-precision CNNs. Our approach incorporates a folding design method to implement shared PEs between layers, significantly reducing FPGA resource usage. Furthermore, we designed a data reading method, incorporating a register set buffer between memory and processing elements to alleviate issues related to mismatched data reading and computing speeds. Our implementation of the mixed-precision ResNet20 model on the Kintex-7 Eco R2 development board achieves an inference accuracy of 91.68% and a computing speed 4.27 times faster than the Central Processing Unit (CPU) on the CIFAR-10 dataset, with an accuracy drop of only 1.21%. Compared to a unified 16-bit FPGA accelerator design method, our proposed approach demonstrates an 89-fold increase in computing speed while maintaining similar accuracy.

show abstract

“…Reducing the model network structure to achieve lightweight is the main means to improve the detection speed, which can be roughly divided into two technical routes: lightweight network design [19] and network pruning [20] according to different implementation methods. Starting from lightweight network design and combining the practical application scenarios of fruit and vegetable recognition, this chapter proposes a classification model based on depth-wise separable neural network, and compares the network model capability with some classical lightweight network models.…”

Section: Fruit and Vegetable Classification Model Based On Depth-wise...mentioning

confidence: 99%

Research on fruit and vegetable classification and recognition method based on Depth-wise Separable Convolution

Yue,

Weng,

Miao

et al. 2023

Preprint

View full text Add to dashboard Cite

Aiming at the problem that the settlement process in agricultural trade markets and fruit and vegetable supermarkets is not intelligent enough and the difficulty of deploying heavy neural network models, the lightweight recognition method of fruit and vegetable classification model is studied. Firstly, aiming at the complex environmental problems of fruit and vegetable supermarkets and farmers' markets and the subsequent practical use, a multi-scene collection scheme was used to collect 170 kinds of fruits and vegetables and 136,000 pictures in the fruit and vegetable supermarket, and an image preprocessing scheme for weakened bagging was formulated to further enhance the data. Then, a fruit and vegetable classification recognition model based on Depth-wise separable convolution is designed, trained and tested. Its top-1 success rate has reached 96.8%, the top-5 success rate has reached 100%. Compared to Mobilenetv2-224, the amount of computation has been reduced by 70%, compared to Mobilenetv3-224, the amount of computing has also been reduced by 60%, and the recognition ability was higher than Mobilenetv2-224 and lower than Mobilenetv3-224. Finally, the problems faced by the designed fruit and vegetable classification model in the actual deployment are analyzed.

show abstract

Dynamic sparsity and model feature learning enhanced training for convolutional neural network-pruning

Cited by 6 publications

References 31 publications

Big data integration technology based Prediction of Sino-Korean cultural products trade

Big data integration technology based Prediction of Sino-Korean cultural products trade

Research on Efficient CNN Acceleration Through Mixed Precision Quantization: A Comprehensive Methodology

Research on fruit and vegetable classification and recognition method based on Depth-wise Separable Convolution

Contact Info

Product

Resources

About