Delving into Semantic Scale Imbalance
Authors: Yanbiao Ma, Licheng Jiao, Fang Liu, Yuxin Li, Shuyuan Yang, Xu Liu
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments show that dynamic semantic-scale-balanced learning consistently enables the model to perform superiorly on large-scale long-tailed and non-long-tailed natural and medical datasets, which is a good starting point for mitigating the prevalent but unnoticed model bias. To validate the superiority and generality of the proposed dynamic semantic-scale-balanced learning, we design four experiments. |
| Researcher Affiliation | Academia | Key Laboratory of Intelligent Perception and Image Understanding of the Ministry of Education School of Artificial Intelligence, Xidian University Xi an, 710071, China {ybmamail,yxli_12}@stu.xidian.edu.cn, lchjiao@mail.xidian.edu.cn, f63liu@163.com, syyang@xidian.edu.cn, xuliu361@163.com |
| Pseudocode | Yes | Algorithm 1 Calculation of Sample Volume and Algorithm 2 Dynamic Re-Weighting Training Framework |
| Open Source Code | No | We will open source the toolkit for measuring the information geometry of data, which includes the application of semantic scale in various scenarios, such as data collection, and representative data selection. |
| Open Datasets | Yes | The sub-datasets generated based on CIFAR-10, CIFAR-100 [30] and Mini-Image Net [62] are shown in Appendix D.1. Image Net-LT and i Naturalist2018 [60] and large-scale Image Net [45] and sample-balanced CIFAR-100, and CUB-2011 [63] and Cars196 [29]. MSCOCO-GLT [56]. The OIA-ODIR dataset [31]. RSSCN7 dataset and NWPU-RESISC45 dataset. |
| Dataset Splits | Yes | We adopt the official training and validation splits [11]. The ILSVRC2012 split contains 1,281,167 training and 50,000 validation images. The CUB-2011 dataset has 5,864 images in the first 100 classes for training and 5,924 images in the second 100 classes for testing. The Cars196 dataset consists of 196 classes totaling 16,185 images, with the first 98 classes for training and the remaining classes for testing. We create three long-tailed CIFAR-100 with the first 60 classes for training (See Figure 8a) and test on the remaining classes [42]. Following the official split, the number of images for training and testing is 50% of the total number each (for RSSCN7). Following the official split, 20% of the images are used for training and 80% for testing (for NWPU-RESISC45). |
| Hardware Specification | Yes | In this work, we use 4 NVIDIA 2080Ti GPUs to train all the models. |
| Software Dependencies | No | The paper mentions software like "Adam optimizer" and implicitly uses frameworks (e.g., PyTorch through code examples like `model(xi)` or `Num Py.mean`), but it does not provide specific version numbers for these software components or libraries required for reproduction. |
| Experiment Setup | Yes | All models are trained using the Adam optimizer [28] with an initial learning rate of 0.01 and then decayed by 0.98 at each epoch. We train Res Net-18 and Res Net-34 on CIFAR-10-LT and CIFAR-10 with an imbalance factor of 200, respectively, and the test set of CIFAR-10-LT is consistent with CIFAR-10. During training, the batch size is fixed to 64, and the optimizer adopts Adam. The learning rate is initially 0.01 and becomes 0.98 the previous learning rate after each epoch. We do not employ other additional tricks and data augmentation strategies. We use Res Next-50 [70] on Image Net LT and Res Net-50 [17] on i Naturalist2018 as the network backbone for all methods. And we conduct model training with the SGD optimizer based on batch size 256 (for Image Net-LT) / 512 (for i Naturalist), momentum 0.9, weight decay factor 0.0005, and learning rate 0.1 (linear LR decay). |