Inverse Weight-Balancing for Deep Long-Tailed Learning
Authors: Wenqi Dang, Zhou Yang, Weisheng Dong, Xin Li, Guangming Shi
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our method can greatly improve performance on imbalanced datasets such as CIFAR100-LT with different imbalance factors, Image Net-LT, and i Naturelists2018. |
| Researcher Affiliation | Academia | 1Xi Dian University, China 2West Virginia University, America 3Peng Cheng Laboratory, China |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a statement or link indicating the availability of open-source code for the described methodology. |
| Open Datasets | Yes | Datasets: We have carried out a series of experiments in CIFAR100-LT (Krizhevsky, Hinton et al. 2009), Image Net LT (Liu et al. 2018), i Naturalist2018 (Van Horn et al. 2018). |
| Dataset Splits | No | The paper mentions that 'the test or valid set is balanced' but does not provide specific percentages or sample counts for a distinct validation set split or refer to a standard validation split. |
| Hardware Specification | Yes | CIFAR100-LT dataset requires only one Ge Force RTX 2080 card, and the other two datasets require 8 Ge Force RTX 2080 cards due to batch size and image size. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers, such as programming language versions or library versions. |
| Experiment Setup | Yes | First, in the first stage, for CIFAR100LT, the batch size is 64, the learning rate is 0.01, and the weight decay is 5e-3. For Image Net-LT, the batch size is 128, the learning rate is 0.01, and the weight decay is 5e-4. For i Naturalist 2018, the batch size is 512, the learning rate is 0.02, and the weight decay is 1e-4. Each of the three datasets trains 200 epochs, and the learning rate uses the cosine decay to 0. Next is the second stage. For CIFAR100-LT, the batch size is 64, the learning rate is 0.005 and the hyperparameter λ=0.15. For Image Net-LT, the batch size is 512, the learning rate is 0.01, and the hyperparameter λ=0.05. For i Naturalist 2018, the batch size is 512, the learning rate is 0.0002, and the hyperparameter λ=0.01. Each of the three datasets trains only 10 epochs in the second stage, and the learning rate uses the cosine decay to 0. |