Effective Data Augmentation With Diffusion Models

Authors: Brandon Trabucco, Kyle Doherty, Max A Gurinas, Ruslan Salakhutdinov

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on few-shot image classification tasks, and on a real-world weed recognition task, and observe an improvement in accuracy in tested domains. Our ablations illustrate that DA-Fusion produces larger gains for the more fine-grain concepts.
Researcher Affiliation Collaboration Brandon Trabucco 1, Kyle Doherty 2, Max Gurinas 3, Ruslan Salakhutdinov 1 1 Carnegie Mellon University, 2 MPG Ranch, 3 University Of Chicago, Laboratory Schools brandon@btrabucco.com, rsalakhu@cs.cmu.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Open-source code is released at: https://github.com/brandontrabucco/da-fusion.
Open Datasets Yes We benchmark our data augmentations on six standard computer vision datasets. We employ Caltech101 (Fei-Fei et al., 2004), Flowers102 (Nilsback and Zisserman, 2008), FGVC Aircraft (Maji et al., 2013), Stanford Cars (Krause et al., 2013), COCO Lin et al. (2014), and PASCAL VOC Everingham et al. (2009). We contribute a dataset of topdown drone images of semi-natural areas in the western United States. These data were gathered in an effort to better map the extent of a problematic invasive plant, leafy spurge (Euphorbia esula)... Additional details are in Appendix H.
Dataset Splits Yes We use the official 2017 training and validation sets of COCO, and the official 2012 training and validation sets of PASCAL VOC. For 15 trials we generated random validation sets with 20 percent of the data, and fine-tuned a pretrained Res Net50 on the remaining 80 percent
Hardware Specification Yes In our experiments, erasing a single class from Stable Diffusion takes two hours on a single 32GB V100 GPU.
Software Dependencies No The paper mentions specific software components like "torchvision" and "Stable Diffusion Checkpoint Comp Vis/stable-diffusion-v1-4", but does not provide specific version numbers for general software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes Specific values for these hyperparameters are given in Table 1.