One Step Learning, One Step Review
Authors: Xiaolong Huang, Qiankun Li, Xueran Li, Xuesong Gao
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments on various tasks such as image classification, object detection, semantic segmentation, and instance segmentation, we demonstrate the general applicability and state-of-the-art performance of our proposed OLOR. |
| Researcher Affiliation | Academia | Xiaolong Huang1, Qiankun Li2, 3*, Xueran Li4, 5, Xuesong Gao1 1School of Artificial Intelligent, Chongqing University of Technology, Chongqing, China 2Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, China 3Department of Automation, University of Science and Technology of China, Hefei, China 4Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei, China 5Anhui University, Hefei, China |
| Pseudocode | Yes | Algorithm 1: OLOR for SGD with Momentum |
| Open Source Code | Yes | Code is available at https://github.com/rainbow-xiao/OLOR-AAAI-2024. |
| Open Datasets | Yes | We experiment on ten popular visual task datasets, i.e., CIFAR-100 (Krizhevsky, Hinton et al. 2009), SVHN (Netzer et al. 2011), CUB-200 (Wah et al. 2011), Stanford Cars (Krause et al. 2013), Places-LT (Zhou et al. 2014), IP102 (Patterson et al. 2014), Office Home (Venkateswara et al. 2017), and PACS (Li et al. 2017), covering general classification, fine-grained classification, long-tailed classification, cross-domain classification, object detection, semantic segmentation, and instance segmentation. More details are listed in Table 1. |
| Dataset Splits | Yes | For training stage, we first pre-train a model using D1 as train set and D2 as valid set for 100 epochs, then finetune the model using D2 as train set and D1 as valid set for 30 epochs through Full and OLOR methods |
| Hardware Specification | Yes | The experiments are performed on two A5000 GPU with 24 GB memory and Ubuntu 20.04 operating system. |
| Software Dependencies | Yes | Python 3.8.3 serves as the programming language, while Py Torch 2.0.0 framework is employed. |
| Experiment Setup | Yes | The input image size is set at 224 224. The batch size varies depending on the freezing strategy. Specifically, 128, 256 and 512 are chosen for full unfreezing, parameter isolated, and full freezing based methods, respectively. Regarding the learning rate, for Conv Ne Xt backbones, we employ the SGD optimizer with a momentum of 0.9. The learning rates differ based on the freezing strategy. In detail, 1e-2, 2e-2 and 4e-2 for full unfreezing, parameter isolated, and full freezing based methods, respectively. For Vi T backbones, we use the Adam optimizer with a momentum of (0.9, 0.999). The learning rates for Vi T backbones also vary according to the freezing strategy, i.e., 1e-4 for full unfreezing, 2e-4 for partial unfreezing, and 4e4 for full freezing. We train on cross-domain datasets for 30 epochs, while for other datasets, we train for 50 epochs. |