Adversarial AutoMixup
Authors: Huafeng Qin, Xin Jin, Yun Jiang, Mounîm El-Yacoubi, Xinbo Gao
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on seven image benchmarks consistently prove that our approach outperforms the state of the art in various classification scenarios. |
| Researcher Affiliation | Academia | 1Chongqing Technology and Business University 2Telecom Sud Paris, Institut Polytechnique de Paris 3Chongqing University of Posts and Telecommunications |
| Pseudocode | Yes | Algorithm 1 Ad Auto Mix training process |
| Open Source Code | Yes | The source code is available at https://github.com/Jin Xins/Adversarial-Auto Mixup. |
| Open Datasets | Yes | (1) CIFAR-100 (Krizhevsky et al., 2009) contains 50,000 training images and 10,000 test images in 32 32 resolutions, with 100 class settings. (2) Tiny-Image Net (Chrabaszcz et al., 2017) contains 10,000 training images and 10,000 validation images of 200 classes in 64 64 resolutions. (3) Image Net-1K (Krizhevsky et al., 2012) contains 1,281,167 training images and 50,000 validation images of 1000 classes. (4) CUB-200-2011 (Wah et al., 2011) contains 11,788 images from 200 wild bird species. FGVC-Aircrafts (Maji et al., 2013) contains 10,000 images of 100 classes of aricrafts and Standford-Cars (Krause et al., 2013) contains 8,144 training images and 8,041 test images of 196 classes. |
| Dataset Splits | Yes | (2) Tiny-Image Net (Chrabaszcz et al., 2017) contains 10,000 training images and 10,000 validation images of 200 classes in 64 64 resolutions. (3) Image Net-1K (Krizhevsky et al., 2012) contains 1,281,167 training images and 50,000 validation images of 1000 classes. |
| Hardware Specification | No | The paper details training configurations and optimizers but does not specify the exact hardware used for experiments, such as GPU models, CPU types, or memory. |
| Software Dependencies | No | The paper mentions implementing the algorithm on 'Open Mixup' and using 'Py Torch pre-trained models', but it does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | For all classification results, we report the mean performance of 3 trials where the median of top-1 test accuracy in the last 10 training epochs is recorded for each trial. Some common parameters follow the experimental settings of Auto Mix and we provide our own hyperparameters in Appendix A.2. The basic learning rate is 0.1, dynamically adjusted by cosine scheduler, SGD (Loshchilov & Hutter, 2016) optimizer with momentum of 0.9, weight decay of 0.0001, batch size of 100. To train Vi T-based models, e.g. Swin-Tiny Transformer and Conv Ne Xt-Tiny, we train them with Adam W (Loshchilov & Hutter, 2019) optimizer with weight decay of 0.05, batch size of 100, 200 epochs. On Tiny-Image Net, except for a learning rate of 0.2 and training over 400 epochs, training settings are similar to the ones used in CIFAR100. On Image Net-1K, we train Res Net18, Res Net34 and Res Net50 for 100 epochs using Py Torch-style setting. The experiments implementation details are provided in Appendix A.3. |