Global Sparse Momentum SGD for Pruning Very Deep Neural Networks

Authors: Xiaohan Ding, guiguang ding, Xiangxin Zhou, Yuchen Guo, Jungong Han, Ji Liu

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate GSM by pruning several common benchmark models on MNIST, CIFAR-10 [29] and Image Net [9], and comparing with the reported results from several recent competitors. For each trial, we start from a well-trained base model and apply GSM training on all the layers simultaneously.
Researcher Affiliation Collaboration 1 Beijing National Research Center for Information Science and Technology (BNRist); School of Software, Tsinghua University, Beijing, China 2 Department of Electronic Engineering, Tsinghua University, Beijing, China 3 Department of Automation, Tsinghua University; Institute for Brain and Cognitive Sciences, Tsinghua University, Beijing, China 4 WMG Data Science, University of Warwick, Coventry, United Kingdom 5 Kwai Seattle AI Lab, Kwai Fe DA Lab, Kwai AI platform
Pseudocode No The paper provides mathematical formulations (e.g., equations 1, 9, 10) for its update rules but does not include a distinct pseudocode or algorithm block.
Open Source Code Yes The codes are available at https://github.com/Ding Xiao H/GSM-SGD.
Open Datasets Yes We evaluate GSM by pruning several common benchmark models on MNIST, CIFAR-10 [29] and Image Net [9]
Dataset Splits Yes After GSM training, we conduct lossless pruning and test on the validation dataset.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1).
Experiment Setup Yes For MNIST: 'We use momentum coefficient β = 0.99 and a batch size of 256. The learning rate schedule is α = 3 10 2, 3 10 3, 3 10 4 for 160, 40 and 40 epochs, respectively.' For CIFAR-10: 'We use β = 0.98, a batch size of 64 and learning rate α = 5 10 3, 5 10 4, 5 10 5 for 400, 100 and 100 epochs, respectively.'