Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution
Authors: Kai Liu, Zhihang Fu, Sheng Jin, Chao Chen, Ze Chen, Rongxin Jiang, Fan Zhou, Yaowu Chen, Jieping Ye
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our theoretically grounded method translates into consistent improvements on the representative CIFAR10-LT, CIFAR100-LT, and Image Net-LT benchmarks against several state-of-the-art OOD detection approaches. Code is available at https://github.com/alibaba/imood. and In this section, we empirically validate the effectiveness of our Im OOD on several representative imbalanced OOD detection benchmarks. The experimental setup is described in Sec. 4.1, based on which extensive experiments and discussions are displayed in Sec. 4.2 and Sec. 4.3. |
| Researcher Affiliation | Collaboration | Kai Liu1,2 , Zhihang Fu2 , Sheng Jin2, Chao Chen2, Ze Chen2, Rongxin Jiang1 , Fan Zhou1, Yaowu Chen1, Jieping Ye2 1Zhejiang University, 2Alibaba Cloud |
| Pseudocode | No | The paper describes its method in text and mathematical formulations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/alibaba/imood. |
| Open Datasets | Yes | Following the literature [51, 23, 10, 40], we use the popular CIFAR10-LT, CIFAR100LT [8], and Image Net-LT [34] as imbalanced in-distribution datasets. |
| Dataset Splits | Yes | The original CIAFR10/100 test sets are kept for evaluating the ID classification capability. For OOD detection, the Tiny Images80M [50] is adopted as the auxiliary OOD training data, and the test set is semantically coherent out-of-distribution (SC-OOD) benchmark [56]. and For the large-scale Image Net-LT benchmark, training samples are sampled from the original Image Net-1k [12] dataset, and the validation set is taken for evaluation. |
| Hardware Specification | Yes | We use 2 NVIDIA V100-32G GPUs in all our experiments. |
| Software Dependencies | No | The paper mentions using Adam optimizer and SGD optimizer, and ResNet models, but does not specify software versions for libraries like PyTorch, TensorFlow, or Python itself. |
| Experiment Setup | Yes | On CIFAR10/100-LT benchmarks, we train ResNet18 [18] for 200 epochs using Adam optimizer, with a batch size of 256. The initial learning rate is 0.001, which is decayed to 0 using a cosine annealing scheduler. The weight decay is 5e-4. On the ImageNet-LT benchmark, we train ResNet50 [18] for 100 epochs with SGD optimizer with the momentum of 0.9. The batch size is 256. The initial learning rate is 0.1, which is decayed by a factor of 10 at epochs 60 and 80. The weight decay is 5e-5. |