Learning Instance-Specific Augmentations by Capturing Local Invariances

Authors: Ning Miao, Tom Rainforth, Emile Mathieu, Yann Dubois, Yee Whye Teh, Adam Foster, Hyunjik Kim

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate that Insta Aug learns meaningful input-dependent augmentations for a wide range of transformation classes, which in turn provides better performance on both supervised and self-supervised tasks.
Researcher Affiliation Collaboration 1Dept. of Statistics, University of Oxford 2Stanford University 3Microsoft Research 4Deep Mind, UK.
Pseudocode Yes Algorithm 1 Location related parameterization
Open Source Code Yes Accompanying code is provided at https://github.com/Ning Miao/Insta Aug.
Open Datasets Yes We first evaluate the performance of jointly training Insta Aug and the classifier on Tiny-Imagenet (Tiny IN, 64 64)... We exploit this on the larger Imagenet dataset (224 224) (Deng et al., 2009)... We benchmark on the texture classification dataset Raw Foo T (Bianco et al., 2017).
Dataset Splits Yes A scheduler is used to decrease the learning rate by a factor of 0.9 once validation accuracy doesn t increase for 10 epochs. ... To further investigate the effect of each augmentation method, we additionally split the 46 test sets into two equally-sized groups.
Hardware Specification Yes On a single 1080Ti, each iteration of training Insta Aug on Tiny IN takes 0.25s
Software Dependencies No The paper mentions using SGD and Adam optimizers, and refers to 'Mixmo codebase' and a codebase from Ermolov et al., 2021, which implies software frameworks like PyTorch. However, it does not specify exact version numbers for any software libraries or dependencies.
Experiment Setup Yes For the classifier, the initial learning rate is set to 0.2 (with momentum 0.9 and weight decay 1e 4). ... The learning rate of the augmentation module ϕ is fixed at 1e 5. Batch size is set to 100 and we pre-train Insta Aug for 10 epochs without augmentation. We train the model until convergence and the maximum epoch is set to 150.