Denoised Maximum Classifier Discrepancy for Source-Free Unsupervised Domain Adaptation

Authors: Tong Chu, Yahao Liu, Jinhong Deng, Wen Li, Lixin Duan472-480

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

Reproducibility Variable Result LLM Response
Research Type Experimental We achieve state-of-the-art results on three domain adaptation benchmark datasets, which clearly validates the effectiveness of our proposed approach. Full code is available at https://github.com/kkkkkkon/D-MCD.
Researcher Affiliation Academia School of Computer Science and Engineering & Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China {uestcchutong, lyhaolive, jhdeng1997, liwenbnu, lxduan}@gmail.com
Pseudocode No The paper describes steps in the training process (e.g., "Training process of D-MCD" and "Model Adaptation" steps A, B, C) but does not present them in a formalized pseudocode or algorithm block.
Open Source Code Yes Full code is available at https://github.com/kkkkkkon/D-MCD.
Open Datasets Yes We evaluate our method on three widely used UDA benchmark datasets: 1) VISDA (Peng et al. 2017), a large-scale challenging dataset with 12 classes; 2) Office Home (Venkateswara et al. 2017), a medium-sized image classification dataset with four distinctive domains (Art (A), Clipart (C), Product (P), and Real World (R)); 3) Office31 (Saenko et al. 2010), a small-sized image classification dataset comprising three different domains (Amazon (A), DSLR (D), and Webcam (W)).
Dataset Splits No The paper mentions using well-known benchmark datasets but does not explicitly state the specific training, validation, and test splits (e.g., percentages or counts) used for reproducibility. While benchmarks often have standard splits, the paper does not specify them.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. It only mentions using ResNet models, which are typically trained on GPUs, but no hardware specifics are given.
Software Dependencies No The paper does not provide specific version numbers for software components or libraries used in the experiments (e.g., Python version, PyTorch version, CUDA version).
Experiment Setup Yes We set the following hyperparameters for RCE loss β = 0.1 for Office and β = 0.001 for VISDA, γ = 0.0025 in training step B and C, r = 0.4 and c = 0.2 for VISDA and Office-Home and r = 0.6, c = 0.1 for Office31. We adopt the Stochastic Gradient Descent optimizer (SGD) with momentum 0.9 and weight decay 5 10 4 and the same learning rate scheduler η = η0 (1 + 10 p) 0.75 where p is the training progress changing from 0 to 1. For the VISDA dataset, the learning rates for the feature extractor and the feature classifier are set to 3 10 4 and 1 10 3 respectively. For the Office-Home and Office31 datasets, learning rate of the feature extractor is 3 10 3 and the learning rate of the feature classifier is 1 10 2.