Transferability vs. Discriminability: Batch Spectral Penalization for Adversarial Domain Adaptation

Authors: Xinyang Chen, Sinan Wang, Mingsheng Long, Jianmin Wang

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that the approach significantly improves upon representative adversarial domain adaptation methods to yield state of the art results. (Abstract) and tables like Table 1. Accuracy (%) on Office-31 for unsupervised domain adaptation (Res Net-50).
Researcher Affiliation Academia Xinyang Chen 1 Sinan Wang 1 Mingsheng Long 1 Jianmin Wang 1 1School of Software, BNRist, Research Center for Big Data, Tsinghua University. E-mail: chenxiny17@mails.tsinghua.edu.cn. Correspondence to: M. Long <mingsheng@tsinghua.edu.cn>.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code of BSP is available at github.com/thuml/ Batch-Spectral-Penalization.
Open Datasets Yes Office-31 (Saenko et al., 2010) is a vanilla dataset for visual domain adaptation with 4,652 images in 31 categories from three domains: Amazon (A), Webcam (W) and DSLR (D). Office-Home (Venkateswara et al., 2017) is a more difficult dataset than Office-31, which consists of around 15,500 images from 65 classes in office and home settings... Vis DA-2017 (Peng et al., 2017) is a challenging simulation-to-real dataset... Digits (Ganin et al., 2016). We use three digits datasets: MNIST, USPS, and SVHN.
Dataset Splits No The paper does not explicitly provide training/test/validation dataset splits or cross-validation setup needed to reproduce the experiment beyond stating that all labeled source and unlabeled target samples participate in training.
Hardware Specification Yes In addition, in each min-batch iteration, (c) uses 0.342s while (d) uses 0.359s (Titan V), which shows that SVD does not cost much computation for feature matrix of a small batch.
Software Dependencies No The paper mentions 'Py Torch' as the implementation framework but does not specify its version number or versions of other key software components.
Experiment Setup Yes We fix δ = 1 and β = 10 4 in all experiments. The learning rates of the layers trained from scratch are set to be 10 times those of fine-tuned layers. We adopt mini-batch SGD with momentum of 0.95 using the learning rate and progressive training strategies of DANN (Ganin & Lempitsky, 2015).