BATUDE: Budget-Aware Neural Network Compression Based on Tucker Decomposition

Yin, Miao; Phan, Huy; Zang, Xiao; Liao, Siyu; Yuan, Bo

doi:10.1609/aaai.v36i8.20869

Cited by 7 publications

(2 citation statements)

References 25 publications

(45 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Automatic Rank Selection. Our work is most closely related to (Gusak et al 2019;Liebenwein et al 2021;Li et al 2022;Yin et al 2022a), the state-of-the-art automatic rank selection solutions. Specifically, (Gusak et al 2019) proposes to utilize variational Bayesian matrix factorization to determine the ranks of the tensor decomposed DNNs in a multi-stage way.…”

Section: Related Workmentioning

confidence: 96%

HALOC: Hardware-Aware Automatic Low-Rank Compression for Compact Neural Networks

Xiao¹,

Zhang²,

Gong³

et al. 2023

Preprint

View full text Add to dashboard Cite

Low-rank compression is an important model compression strategy for obtaining compact neural network models. In general, because the rank values directly determine the model complexity and model accuracy, proper selection of layerwise rank is very critical and desired. To date, though many low-rank compression approaches, either selecting the ranks in a manual or automatic way, have been proposed, they suffer from costly manual trials or unsatisfied compression performance. In addition, all of the existing works are not designed in a hardware-aware way, limiting the practical performance of the compressed models on real-world hardware platforms. To address these challenges, in this paper we propose HALOC, a hardware-aware automatic low-rank compression framework. By interpreting automatic rank selection from an architecture search perspective, we develop an end-to-end solution to determine the suitable layer-wise ranks in a differentiable and hardware-aware way. We further propose design principles and mitigation strategy to efficiently explore the rank space and reduce the potential interference problem. Experimental results on different datasets and hardware platforms demonstrate the effectiveness of our proposed approach. On CIFAR-10 dataset, HALOC enables 0.07% and 0.38% accuracy increase over the uncompressed ResNet-20 and VGG-16 models with 72.20% and 86.44% fewer FLOPs, respectively. On ImageNet dataset, HALOC achieves 0.9% higher top-1 accuracy than the original ResNet-18 model with 66.16% fewer FLOPs. HALOC also shows 0.66% higher top-1 accuracy increase than the state-of-the-art automatic low-rank compression solution with fewer computational and memory costs. In addition, HALOC demonstrates the practical speedups on different hardware platforms, verified by the measurement results on desktop GPU, embedded GPU and ASIC accelerator.

show abstract

Section: Related Workmentioning

confidence: 96%

HALOC: Hardware-Aware Automatic Low-Rank Compression for Compact Neural Networks

Xiao¹,

Zhang²,

Gong³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Miao et al [36] recently proposed a budget-aware rank selection method that can calculate tensor ranks via one-shot training. Although it can automatically select the proper tensor ranks for each layer, it may obtain different ranks when the training environment changes, such as the dataset and training hyperparameters.…”

Section: Related Workmentioning

confidence: 99%

Remote Sensing Imagery Object Detection Model Compression via Tucker Decomposition

Huyan

Jiang

et al. 2023

Mathematics

View full text Add to dashboard Cite

Although convolutional neural networks (CNNs) have made significant progress, their deployment onboard is still challenging because of their complexity and high processing cost. Tensors provide a natural and compact representation of CNN weights via suitable low-rank approximations. A novel decomposed module called DecomResnet based on Tucker decomposition was proposed to deploy a CNN object detection model on a satellite. We proposed a remote sensing image object detection model compression framework based on low-rank decomposition which consisted of four steps, namely (1) model initialization, (2) initial training, (3) decomposition of the trained model and reconstruction of the decomposed model, and (4) fine-tuning. To validate the performance of the decomposed model in our real mission, we constructed a dataset containing only two classes of objects based on the DOTA and HRSC2016. The proposed method was comprehensively evaluated on the NWPU VHR-10 dataset and the CAST-RS2 dataset created in this work. The experimental results demonstrated that the proposed method, which was based on Resnet-50, could achieve up to 4.44 times the compression ratio and 5.71 times the speedup ratio with merely a 1.9% decrease in the mAP (mean average precision) of the CAST-RS2 dataset and a 5.3% decrease the mAP of the NWPU VHR-10 dataset.

show abstract