Bidirectional Adaptation for Robust Semi-Supervised Learning with Inconsistent Data Distributions
Authors: Lin-Han Jia, Lan-Zhe Guo, Zhi Zhou, Jie-Jing Shao, Yuke Xiang, Yu-Feng Li
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results demonstrate the effectiveness of our proposed framework. We conduct extensive experiments on image datasets to demonstrate the effectiveness of our proposed method. |
| Researcher Affiliation | Collaboration | 1National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China 2Consumer BG, Huawei Technologies, Shenzhen, China. Correspondence to: Yu-Feng Li <liyf@lamda.nju.edu.cn>. |
| Pseudocode | Yes | Algorithm 1 Bidirectional Adaptation. Input: labeled dataset DL = {(xl 1, y1), . . . , (xl nl, ynl)}, unlabeled dataset DU = {xu 1, . . . , xu nu}, DA algorithm A, the total number of iterations T, the learning rate η, the batch size B, the number of classes k, the ratio of the number of labeled samples to the number of unlabeled samples µ in each batch, the ratio of unsupervised loss to supervised loss λu. Output: pseudo-label predictor h, target predictor f. |
| Open Source Code | No | The algorithm implementations are based on the SSL toolkit LAMDA-SSL (Jia et al., 2023). |
| Open Datasets | Yes | We conduct an experiment on the extracted features of Office-Caltech (Gong et al., 2012) dataset to prove that both of the distribution discrepancies we explain in our theoretical framework actually exist, and only by alleviating both can the performance be the best. We selected three commonly used datasets with multiple domains: Image-CLEF (Caputo et al., 2014), Office31 (Saenko et al., 2010), and Vis DA-2017 (Peng et al., 2018). |
| Dataset Splits | No | No explicit validation split information was provided beyond the common practice of using 'test' and general data splits. |
| Hardware Specification | Yes | All experiments in Section 6.1 are conducted with a single Intel(R) Core(TM) i7-9750H CPU. All experiments in Section 6.2 and Appendix B are conducted with 4 NVIDIA Ge Force RTX 3090 GPUs and 12 NVIDIA Tesla V100 GPUs. This work uses the Huawei Mind Spore platform for experimental testing partially. |
| Software Dependencies | No | The paper mentions using "Py Torch" for implementation and indicates that "The algorithm implementations are based on the SSL toolkit LAMDA-SSL (Jia et al., 2023)". It also mentions "scikit-learn" but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | The batch size B is set to 64, the max iteration T is set to 2000, the ratio of unlabeled to labeled data µ is set to 1.0, and the ratio of unsupervised loss λu is set to 0.1. For fairness, all methods use Res Net-50 (He et al., 2016) pre-trained on Image Net (Russakovsky et al., 2015) as the backbone network and SGD with the initial learning rate 5 10 4 as the optimizer. |