Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
[Re] CUDA: Curriculum of Data Augmentation for Long‐tailed Recognition
Authors: Barath Chandran.C
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this reproducibility study, we present our results and experience while replicating the paper titled CUDA: Curriculum of Data Augmentation for Long-Tailed Recognition (Ahn et al., 2023). ... We successfully replicated a substantial portion of the results pertaining to the long-tailed CIFAR-100-LT dataset and extended our analysis to provide deeper insights into how CUDA efficiently tackles class imbalance. |
| Researcher Affiliation | Academia | Barath Chandran.C EMAIL Indian Institute of Technology Roorkee |
| Pseudocode | No | The paper describes the methodology for CUDA and its components (strength-based augmentation and updating Level-of-Learning score) in textual paragraphs and mathematical formulas, but it does not present any clearly labeled 'Algorithm' or 'Pseudocode' block. |
| Open Source Code | Yes | The code and the readings are available here. |
| Open Datasets | Yes | The reproduction study is done on the dataset CIFAR-100-LT as mentioned in the previous works(Cao et al., 2019). The CIFAR-100 dataset (Canadian Institute for Advanced Research, 100 classes) is a subset of the Tiny Images dataset and consists of 60000 32x32 color images. |
| Dataset Splits | Yes | The reproduction study is done on the dataset CIFAR-100-LT as mentioned in the previous works(Cao et al., 2019). ... Three different datasets are derived from the CIFAR-100 dataset with imbalance ratios 100, 50, and 10, where the imbalance ratio is defined as |D1| /|D100| =Nmax/Nmin. ... We measure the validation accuracy of CUDA when used with CE (Cross-Entropy), CE-DRW (Cross entropy Dynamic reweighting) (Cao et al., 2019), LDAM-DRW (label-distribution-aware margin loss) (Cao et al., 2019), BS (balanced softmax) (Ren et al., 2020), and RIDE (Wang et al., 2021) for the CIFAR-100-LT dataset following the general settings outlined in Cao et al. (2019). |
| Hardware Specification | Yes | We trained the model in Kaggle with 1 NVIDIA Tesla P100 as the GPU accelerator. |
| Software Dependencies | No | Dependencies were not explicitly provided, and the versions of different libraries were determined through trial and error. |
| Experiment Setup | Yes | The hyperparameters augmentation probability, number of test samples, and acceptance rate are set according to the values stated in the original paper: 0.6 for augmentation probability, 10 for the number of test samples, and 0.5 for the acceptance rate. ... The network is trained using stochastic gradient descent (SGD) with a momentum of 0.9 and a weight decay of 0.0002. The initial learning rate is set to 0.1, and a linear learning rate warm-up is applied during the first 5 epochs to reach the initial learning rate. The training process spans over 200 epochs, during which the learning rate is decayed at the 160th and 180th epochs by a factor of 0.01. ... The average training time of the model was approximately 40 minutes with a batch size of 128 for 200 epochs. |