DropCov: A Simple yet Effective Method for Improving Deep Architectures

Authors: Qilong Wang, Mingze Gao, Zhaolin Zhang, Jiangtao Xie, Peihua Li, Qinghua Hu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various benchmarks (i.e., Image Net-1K, Image Net-C, Image Net-A, Stylized-Image Net, and i Nat2017) show our Drop Cov is superior to the counterparts in terms of efficiency and effectiveness, and provides a simple yet effective method to improve performance of deep architectures involving both deep convolutional neural networks (CNNs) and vision transformers (Vi Ts).
Researcher Affiliation Academia Qilong Wang1,3, Mingze Gao1, Zhaolin Zhang1, Jiangtao Xie2, Peihua Li2, Qinghua Hu1,3, 1Tianjin University, China, 2Dalian University of Technology, China, 3 Haihe Laboratory of Information Technology Application Innovation, Tianjin, China
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No Source code (e.g., Py Torch [33], Paddle Paddle and Mindspore) will be available at https://github.com/mingze G/Drop Cov.
Open Datasets Yes Extensive experiments on various benchmarks (i.e., Image Net-1K, Image Net-C, Image Net-A, Stylized-Image Net, and i Nat2017) show our Drop Cov is superior to the counterparts in terms of efficiency and effectiveness, and provides a simple yet effective method to improve performance of deep architectures involving both deep convolutional neural networks (CNNs) and vision transformers (Vi Ts).
Dataset Splits Yes All models are optimized using the same hyper-parameter settings and data augmentations as suggested in [17, 29], where stochastic gradient descent (SGD) with initial learning rate (lr) of 0.1 is used to train the networks within 100 epochs and lr is decayed by 10 every 30 epochs. We conduct experiments on Image Net-1K (IN-1K) using backbone of Res Net-18.
Hardware Specification Yes All programs run a sever equipped with eight Nvidia RTX 3090 GPUs and 128G RAM.
Software Dependencies No Source code (e.g., Py Torch [33], Paddle Paddle and Mindspore) will be available at https://github.com/mingze G/Drop Cov. The paper lists software names but does not provide specific version numbers for these dependencies.
Experiment Setup Yes All models are optimized using the same hyper-parameter settings and data augmentations as suggested in [17, 29], where stochastic gradient descent (SGD) with initial learning rate (lr) of 0.1 is used to train the networks within 100 epochs and lr is decayed by 10 every 30 epochs.