L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer

Authors: Jin Shin, Hyun Kim

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed method integrated into Res Net-26 and Res Net-50 models demonstrates its robustness by achieving outstanding TTA performance while using the least amount of memory compared to existing studies on CIFAR-10-C, Image Net-C, and Cityscapes-C benchmark datasets.
Researcher Affiliation Academia Jin Shin, Hyun Kim Department of Electrical and Information Engineering and RCEIT Seoul National University of Science and Technology Seoul, Korea {shinjin0103, hyunkim}@seoultech.ac.kr
Pseudocode No The paper includes diagrams (e.g., Figure 2) and mathematical equations, but it does not feature a dedicated 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes The code is available at https://github.com/janus103/L_TTA.
Open Datasets Yes The proposed method integrated into Res Net-26 and Res Net-50 models demonstrates its robustness by achieving outstanding TTA performance while using the least amount of memory compared to existing studies on CIFAR-10-C, Image Net-C, and Cityscapes-C benchmark datasets. To evaluate the robustness of the model in the existing TTA setting, we use the benchmark dataset with the well-established 15 types of corruptions (e.g., noise, blur, weather, and digital) [22]. These are called CIFAR-10-C, CIFAR-100-C, and Image Net-C, respectively.
Dataset Splits Yes Each corruption is applied to the validation set of the original dataset and has identical content information. These are called CIFAR-10-C, CIFAR-100-C, and Image Net-C, respectively. The severity of the corruption is divided into five levels.
Hardware Specification Yes All experiments were conducted on a system equipped with an Intel Xeon Gold 5218R CPU and an NVIDIA Tesla A100 80G GPU.
Software Dependencies No The paper mentions 'pytorch code as supplementary materials' and refers to standard optimizers like 'SGD [1]', but does not provide specific version numbers for software libraries or dependencies.
Experiment Setup Yes In both TTA and warm-up settings, the batch size and learning rate are set to 128 and 0.05, respectively. For CIFAR-10 and CIFAR-100, the model utilizes weights pre-trained on Image Net and is trained for 150 epochs, using the same batch size and learning rate configuration as in the Image Net training setup. The optimizer used is standard SGD [1], consistent across all datasets.