You Only Cut Once: Boosting Data Augmentation with a Single Cut

Authors: Junlin Han, Pengfei Fang, Weihao Li, Jie Hong, Mohammad Ali Armin, Ian Reid, Lars Petersson, Hongdong Li

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Thorough experiments are conducted to evaluate its effectiveness. We first demonstrate that YOCO can be seamlessly applied to varying data augmentations, neural network architectures, and brings performance gains on CIFAR and Image Net classification tasks, sometimes surpassing conventional image-level augmentation by large margins.
Researcher Affiliation Collaboration 1Data61-CSIRO, Canberra, Auistralia 2Australian National University, Canberra, Australia 3University of Adelaide, Adelaide, Australia.
Pseudocode No No structured pseudocode or algorithm blocks were found.
Open Source Code Yes Code is available at Git Hub.
Open Datasets Yes For image classification tasks on CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009) datasets...Next, we validate YOCO on Image Net-1K (henceforth referred to as Image Net) (Russakovsky et al., 2015).
Dataset Splits No No explicit mention of specific validation dataset splits (percentages, counts, or explicit "validation set" statements) was found. The paper describes training and testing protocols for various datasets: "We train 5 CNN architectures (Pre Res Net18 (He et al., 2016a), Xception (Chollet, 2017), Dense Net121 (Huang et al., 2017), Res Ne Xt50 (Xie et al., 2017), WRN-28-10 (Zagoruyko & Komodakis, 2016)) and 2 Vi T models (Vi T (Dosovitskiy et al., 2020) and Swin (Liu et al., 2021)) under the same training recipe, which mostly follows the training setting used in Co-Mixup (Kim et al., 2021)."
Hardware Specification No No specific hardware models (e.g., GPU/CPU models, processors, memory details) were provided. The paper only mentions the number of GPUs used: "models are trained on 2 GPUs" and "4 GPUs are used for training".
Software Dependencies No No specific version numbers for software dependencies were provided. The paper mentions using "Torchattacks (Kim, 2020)", "timm (Wightman, 2019)", and "Pytorch/Torchvision implementation" without version numbers.
Experiment Setup Yes That is, models are trained for 300 epochs, with an initial learning rate of 0.1 decayed by a factor 0.1 at epochs 100 and 200. We employ the SGD optimizer with a momentum of 0.9 and a weight decay of 0.0001. We use a batch size of 100.