This data repository includes the features and the trained backbone parameters used in the ICLR 2022 Paper "On the Importance of Firth Bias Reduction in Few-Shot Classification".
The code accompanying this data is open-source and available at https://github.com/ehsansaleh/firth_bias_reduction
The code and the data have three modules:
1. The "code_firth" module (10 files) relates to the basic ResNet backbones and logistic classifiers (e.g., Figures 2 and 3 in the main paper).
2. The "code_s2m2rf" module (2 files) relates to the S2M2R feature backbones and cosine classifiers (e.g., Figure 4 in the main paper).
3. The "code_dcf" module (3 files) relates to the few-shot Distribution Calibration (DC) method (e.g., Table 1 in the main paper).
The relevant files for each module have the module name as a prefix in their name.
1. For instance, the "code_dcf_features.tar" file should be placed at the "features" directory of the "code_dcf" module.
2. As another example, "code_firth_features_cifarfs_novel.tar" should be placed in the "features" directory of the "code_firth" module, and it includes the features extracted from the novel split of mini-ImageNet dataset.
Each tar-ball should be extracted in its relevant directory, and the md5 check-sums of the extracted files are also provided in the open-source code repository for verification.
Please note that the actual datasets of images are not included here (since we do not own those datasets). However, helper scripts for automatically downloading the original datasets are also provided in the every module and sub-directory of the GitHub code repository.
- Few-Shot Learning
- Firth Bias Reduction
- Few-Shot Classification
- Computer Vision