Domain Adaptation with Adversarial Training on Penultimate Activations

Authors: Tao Sun, Cheng Lu, Haibin Ling

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted on popular UDA benchmarks under both standard setting and source-data free setting. The results validate that our method achieves the best scores against previous arts.
Researcher Affiliation Collaboration Tao Sun 1, Cheng Lu 2, Haibin Ling 1 1 Stony Brook University, USA 2 XPeng Motors, USA
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https: //github.com/tsun/APA.
Open Datasets Yes Datasets. Office-Home (OH) has 65 classes from four domains: Artistic (A), Clip Art (C), Product (P), and Realworld (R). We use both the original version and the RS-UT (Reverse-unbalanced Source and Unbalanced Target) version (Tan, Peng, and Saenko 2020) that is manually created to have a large label shift. Vis DA-2017 (Peng et al. 2017) is a synthetic-to-real dataset of 12 objects. Domain Net (Peng et al. 2019) (DN) is a large UDA benchmark. We use the 40-class version (Tan, Peng, and Saenko 2020) from four domains: Clipart (C), Painting (P), Real (R), Sketch (S).
Dataset Splits No The paper describes the use of source and target domains and discusses training parameters, but it does not explicitly state specific train/validation/test dataset splits (e.g., percentages or sample counts) needed for reproduction.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper states 'We implement our methods with Py Torch.' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes For all tasks, we use batch size 16, β = 0.1, τ = 0.75, ϵ = 30 for APAu, ϵ = 1.0 for APAn, with an only exception on Vis DA where we use β = 0.04 for APAn instead.