A Group-Theoretic Framework for Data Augmentation
Authors: Shuxiao Chen, Edgar Dobriban, Jane H. Lee
JMLR 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | See Figure 1 for a small experiment (see Appendix D for details). ...Figure 1: Benefits of data augmentation: A comparison of test accuracy across training epochs of Res Net18 (He et al., 2016) (1) without data augmentation, (2) horizontally flipping the image with 0.5 probability, and (3) a composition of randomly cropping a 32 32 portion of the image and random horizontal flip. The experiment is repeated 15 times, with the dotted lines showing the average test accuracy and the shaded regions representing 1 standard deviation around the mean. |
| Researcher Affiliation | Academia | Shuxiao Chen EMAIL Edgar Dobriban EMAIL Department of Statistics The Wharton School of the University of Pennsylvania Philadelphia, PA, 19104-6340, USA Jane H. Lee EMAIL Department of Mathematics, and Computer and Information Science University of Pennsylvania Philadelphia, PA 19104-6309, USA |
| Pseudocode | Yes | Algorithm 1: Augmented SGD Input : Data Xi, i = 1, . . . , n; Method to compute gradients L(θ, X) of the loss; Method to sample augmentations g G, g Q; Learning rates ηt; Batch sizes |St|; Initial parameters θ0; Stopping criterion. Output: Final parameters. Set t = 0 While stopping criterion is not met Sample random minibatch St {1, . . . , n} Sample random augmentation gi,t Q for each batch element Update parameters i St L(θ, gi,t Xi) t t + 1 return θ |
| Open Source Code | Yes | Our code is available at https://github.com/dobriban/data_aug. |
| Open Datasets | Yes | We train Res Net18 (He et al., 2016) on CIFAR10 (Krizhevsky, 2009)... The CIFAR10 dataset is standard and can be downloaded from https://www.cs.toronto. edu/~kriz/cifar.html. |
| Dataset Splits | Yes | The left graph shows results from training on the full CIFAR10 training data and the right uses half of the training data as that of the left. |
| Hardware Specification | Yes | This experiment was done on a p3.2xlarge (GPU) instance on Amazon Web Services (AWS). |
| Software Dependencies | No | The paper mentions software components like "Res Net18" and "pytorch-cifar" but does not specify version numbers for these or other software libraries used. |
| Experiment Setup | Yes | We use the default settings from that code, including the SGD optimizer with a learning rate of 0.1, momentum 0.9, weight decay 5 10 4, and batch size of 128. |