Dynamic Sparsity Is Channel-Level Sparsity Learner

Authors: Lu Yin, Gen Li, Meng Fang, Li Shen, Tianjin Huang, Zhangyang "Atlas" Wang, Vlado Menkovski, Xiaolong Ma, Mykola Pechenizkiy, Shiwei Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results demonstrate that Chase achieves 1.7 inference throughput speedup on common GPU devices without compromising accuracy with Res Net-50 on Image Net. We release our codes in https://github.com/luuyin/chase.
Researcher Affiliation Collaboration 1Eindhoven University of Technology, 2Clemson University 3University of Liverpool, 4JD Explore Academy, 5University of Texas at Austin
Pseudocode Yes Algorithm 1: Pseudocode of Chase. Algorithm 2: Overview of Global Parameter Exploration
Open Source Code Yes We release our codes in https://github.com/luuyin/chase.
Open Datasets Yes Our evaluation is conducted with two widely used model architectures VGG-19 [48] and Res Net-50 [15] on across various datasets including CIFAR-10/100 and Image Net
Dataset Splits Yes Our evaluation is conducted with two widely used model architectures VGG-19 [48] and Res Net-50 [15] on across various datasets including CIFAR-10/100 and Image Net
Hardware Specification Yes All results are averaged from 100 individual runs with one NVIDIA 2080TI GPU in float32 on Py Torch. We set the batch size to 128 for CIFAR-100 and 2 for Image Net, when evaluating the latency.
Software Dependencies No The paper mentions "Py Torch" but does not specify its version number. No other specific software dependencies with version numbers are provided.
Experiment Setup Yes Table 9: Implementation hyperparameters of Chase in Table 3, on CIFAR-10/100. Table 10: Implementation hyperparameters of Chase in Table 4, on Image Net. These tables detail hyperparameters such as τtotal (epochs), τstop (epochs), T (iterations), Tp (iterations), β, BS (batch size), LR (learning rate), LR Drop, Optimizer, Momentum, WD (weight decay), and Sparse Init.