Scaling of Class-wise Training Losses for Post-hoc Calibration

Authors: Seungjin Jung, Seungmo Seo, Yonghyun Jeong, Jongwon Choi

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the proposed framework by employing it in the various post-hoc calibration methods, which generally improves calibration performance while preserving accuracy, and discover through the investigation that our approach performs well with unbalanced datasets and untuned hyperparameters.
Researcher Affiliation Collaboration 1Department of Artificial Intelligence, Chung-Ang University, Seoul, Korea 2Naver CLOVA, Seongnam, Korea 3Department of Advanced Imaging, Chung-Ang University, Seoul, Korea.
Pseudocode No The paper describes its methods using prose and mathematical equations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available online2. 2https://github.com/SeungjinJung/SCTL
Open Datasets Yes We use three different datasets and six different pre-trained models to train and evaluate the calibration methods. We separate the datasets into validation datasets and test datasets with the sizes of 25000/10000 for CIFAR10 and CIFAR100 datasets (Krizhevsky & Hinton, 2009) and 25000/25000 for Image Net dataset (Deng et al., 2009).
Dataset Splits Yes We separate the datasets into validation datasets and test datasets with the sizes of 25000/10000 for CIFAR10 and CIFAR100 datasets (Krizhevsky & Hinton, 2009) and 25000/25000 for Image Net dataset (Deng et al., 2009).
Hardware Specification Yes We conduct our experiments upon one RTX3090 environment.
Software Dependencies No The paper mentions optimizers like LBFGS and Adam but does not provide specific version numbers for any software dependencies or libraries used in the implementation.
Experiment Setup Yes We train baseline methods using a learning rate of 0.02, 1000 epochs, and a cross-entropy loss. TS, ETS, and CTS use LBFGS optimizer... and PTS utilizes Adam optimizer with 0.002 weight decay. We initialize α and β by 1.0 and 1.5, respectively, before their optimization. In the learning process, we use the hyperparameters referred to in Table. 4.