DecAug: Out-of-Distribution Generalization via Decomposed Feature Representation and Semantic Augmentation
Authors: Haoyue Bai, Rui Sun, Lanqing Hong, Fengwei Zhou, Nanyang Ye, Han-Jia Ye, S.-H. Gary Chan, Zhenguo Li6705-6713
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that Dec Aug outperforms other state-of-the-art methods on various Oo D datasets, which is among the very few methods that can deal with different types of Oo D generalization challenges. |
| Researcher Affiliation | Collaboration | 1 The Hong Kong University of Science and Technology 2 Huawei Noah s Ark Lab 3 Shanghai Jiao Tong University 4 Nanjing University |
| Pseudocode | Yes | Algorithm 1 Dec Aug: Decomposed Feature Representation and Semantic Augmentation for Oo D generalization |
| Open Source Code | No | The paper does not provide an explicit statement about the open-source availability of their code or a direct link to a repository. |
| Open Datasets | Yes | We evaluate our method on three challenging Oo D datasets with different levels of correlation shift and diversity shift as discussed above: Colored MNIST (Arjovsky et al. 2019), PACS (Li et al. 2017a), and NICO (He, Shen, and Cui 2020). |
| Dataset Splits | Yes | We followed the same experimental protocol as in IRM (Arjovsky et al. 2019) on the Colored MNIST dataset... The digits were colored either red or green based on different correlation with the labels to construct different environments (e.g., 80% and 90% for the training environments and 10% for the test environment). For PACS, We followed the same leave-one-domain-out validation experimental protocol as in (Li et al. 2017a)... We followed the same training, validation and test split as in Ji Gen (Carlucci et al. 2019). |
| Hardware Specification | Yes | We conducted experiments on NVIDIA Tesla V100. |
| Software Dependencies | Yes | Our framework was implemented with Py Torch 1.1.0, CUDA v9.0. |
| Experiment Setup | Yes | The backbone network was a three-layer MLP. The total training epoch was 500 and the batch size was the whole training data. We used the SGD optimizer with an initial learning rate of 0.1. The backbone network we used on the PACS dataset was Res Net-18... The number of training epochs was 100. The batch size was 64. We used the SGD optimizer with a learning rate of 0.02. The backbone network was Res Net-18 without pretraining on the NICO dataset. The number of training epochs was 500 and the batch size was 128. We used the SGD optimizer with a learning rate of 0.05. |