Adversarial Teacher-Student Representation Learning for Domain Generalization

Authors: Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, Yu-Chiang Frank Wang

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive image classification experiments on benchmark datasets in multiple and single source DG settings confirm that, our model exhibits sufficient generalization ability and performs favorably against state-of-the-art DG methods.
Researcher Affiliation Collaboration Fu-En Yang1,2 Yuan-Chia Cheng1 Zu-Yun Shiau1 Yu-Chiang Frank Wang1,2 1Graduate Institute of Communication Engineering, National Taiwan University, Taiwan 2ASUS Intelligent Cloud Services, Taiwan
Pseudocode Yes The pseudo code of our Adversarial Teacher-Student Representation Learning is summarized in Algorithm 1 in supplementary material.
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of the methodology described within the paper.
Open Datasets Yes We evaluate our method on several public benchmark datasets. PACS [35] is composed of four data domains... Office-Home [36] is comprised of four domains... Domain Net [37] is a recently proposed large-scale dataset... Due to page limitation, we additionally provide quantitative comparisons on VLCS [39] and Digit-DG [7] datasets in the supplementary material.
Dataset Splits Yes images from source domains are divided into the training split and the validation split, at a ratio of about 9:1.
Hardware Specification Yes In all our experiments, we implement our model using Py Torch and Dassl.pytorch [43] toolbox, and conduct training on a single NVIDIA TESLA V100 GPU with 32 GB memory.
Software Dependencies No The paper mentions using 'Py Torch and Dassl.pytorch [43] toolbox' but does not specify their version numbers, which are necessary for full reproducibility.
Experiment Setup Yes FS is trained with the SGD optimizer, with an initial learning rate of 0.0005, and a batch size of 32 for 60 epochs. The learning rate is decayed by 0.1 after 30 epochs. FT is updated via EMA with the momentum coefficient τ of 0.999 by default.