Channel Importance Matters in Few-Shot Image Classification
Authors: Xu Luo, Jing Xu, Zenglin Xu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Table 1 shows the performance gains brought by this transformation on 5-way 5-shot FSL tasks. We test the transformation on representations trained with different algorithms, including (1) the conventional training methods including cross-entropy (CE) and the S2M2 algorithm (Mangla et al., 2020), (2) meta-learning methods including Proto Net (Snell et al., 2017) (PN), Meta-baseline (Chen et al., 2021) and Meta Opt (Lee et al., 2019), and (3) Mo Co-v2 (He et al., 2020), a unsupervised contrastive learning method. We test these methods with various backbone networks: Conv4 (Vinyals et al., 2016) and four variants of Res Net (He et al., 2016) including Res Net-12 (Oreshkin et al., 2018), WRN28-10 (Zagoruyko & Komodakis, 2016), Res Net-50 and SE-Res Net50 (Hu et al., 2018). |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China 2Harbin Institute of Technology Shenzhen 3Pengcheng Laboratory. Correspondence to: Xu Luo <frank.luox@outlook.com>, Zenglin Xu <xuzenglin@hit.edu.cn>. |
| Pseudocode | No | The paper includes mathematical equations for the transformation function but no distinct section or figure labeled 'Pseudocode' or 'Algorithm', nor any structured steps for a method formatted like code. |
| Open Source Code | No | The paper does not contain any explicit statement about making the source code for their methodology publicly available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | For Dtrain, we choose (1) the train split of mini Image Net (Vinyals et al., 2016) that contains 38400 images from 64 classes; (2) the train split of Image Net 1K (Russakovsky et al., 2015) containing more than 1M images from 1000 classes; (3) train+val split of i Naturalist 2018 (Horn et al., 2018), a finegrained dataset of plants and animals with a total of more than 450000 training images from 8142 classes. |
| Dataset Splits | Yes | For Dtrain, we choose (1) the train split of mini Image Net (Vinyals et al., 2016) that contains 38400 images from 64 classes; (2) the train split of Image Net 1K (Russakovsky et al., 2015) containing more than 1M images from 1000 classes; (3) train+val split of i Naturalist 2018 (Horn et al., 2018), a finegrained dataset of plants and animals with a total of more than 450000 training images from 8142 classes. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions using 'scikit-learn' for Logistic Regression implementation, but it does not specify a version number for this or any other software dependency. |
| Experiment Setup | Yes | For S2M2 and Mo Co-v2 in Table 1, we directly use the official publicly-available pre-trained checkpoints. All other algorithms in Table 1 are trained using a learning rate 0.1 with cosine decay schedule without restart. SGD with momentum 0.9 is adopted as the optimizer. For all meta-learning algorithms, a total of 60000 5-way 5-shot tasks are sampled for training, each of which contains 15 query images per class. The batch size (number of sampled tasks of each iteration) is 4. All other hyperparameters of Meta Opt match the default settings in the original paper. All conventionally-trained algorithms are trained for 60 epochs, and the batch size is set to 128. |