Switch to Generalize: Domain-Switch Learning for Cross-Domain Few-Shot Classification
Authors: Zhengdong Hu, Yifan Sun, Yi Yang
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results confirm that the cross-domain generalization capacity can be inherited from the training stage to the testing stage, validating our key hypothesis. Consequentially, DSL significantly improves cross-domain few-shot classification and sets up new state of the art. |
| Researcher Affiliation | Collaboration | Zhengdong Hu1,2 , Yifan Sun2, Yi Yang3 1 State Key Laboratory of Industrial Control Technology, Zhejiang University 2 Baidu Research, China 3 CCAI, College of Computer Science and Technology, Zhejiang University |
| Pseudocode | No | The paper describes the methodology using text, figures (like Figure 1 and 2), and mathematical equations, but it does not include explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states: "The DSL is reproducible. In the main text, we describe the utilized datasets in DSL... We provide the details about the experimental implementation, the proof of the proposed remark and the analysis of some hyper parameters in appendix." However, it does not provide an explicit statement about releasing code or a link to a code repository. |
| Open Datasets | Yes | We evaluate the proposed DSL on five datasets: i.e., 4 fine-grained datasets (CUB Welinder et al. (2010), Cars Krause et al. (2013), Places Zhou et al. (2018), and Plantae Horn et al. (2018)) and 1 popular large scale dataset mini-Image Net Deng et al. (2009). |
| Dataset Splits | Yes | We adopt the leave-one-out setting, i.e., one of the fine-grained datasets is chosen as a testing set and the other 3 datasets (out of the 4 datasets) are used as the training sets. |
| 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. |
| Software Dependencies | No | The paper mentions using specific optimizers like "Adam Kingma & Ba (2015)" and "SGD Qian (1999)" and the "Res Net-10 He et al. (2016)" backbone, but it does not specify software versions for programming languages, libraries (e.g., PyTorch, TensorFlow), or other dependencies. |
| Experiment Setup | Yes | We train the model with 400 epochs and set the initial learning rate as 1e-3. In each training iteration, the mini-batch size of switchable domain and the basic training set are both 64. In each mini-batch , we randomly sample 16 classes from each domain (4 images per class)... We use the SGD Qian (1999) to optimize the linear classifier layer and the initial learning rate is 0.01. The weight decay of SGD is 0.001 and the SGD momentum is 0.9. ... We set λ = 0.4 as the weight factor. ... We set α = 0.7 as the weight factor. ... Therefore, we set µ = 1.25 as the optimized threshold for domain-general teacher. |