Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization
Authors: Yongyi Su, Xun Xu, Kui Jia
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate TRIBE on four datasets representing real-world TTA settings. TRIBE consistently achieves the state-of-the-art performance across multiple evaluation protocols. |
| Researcher Affiliation | Academia | 1South China University of Technology 2Institute for Infocomm Research, A*STAR 3School of Data Science, The Chinese University of Hong Kong, Shenzhen |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. It provides mathematical equations and descriptions of methods. |
| Open Source Code | Yes | The code is available at https://github.com/Gorilla-Lab-SCUT/TRIBE. |
| Open Datasets | Yes | We evaluate on four test-time adaptation datasets, including CIFAR10-C (Hendrycks and Dietterich 2019), CIFAR100-C (Hendrycks and Dietterich 2019), Image Net C (Hendrycks and Dietterich 2019) and MNIST-C (Mu and Gilmer 2019). |
| Dataset Splits | No | The paper describes using CIFAR10-C, CIFAR100-C, Image Net C, and MNIST-C for evaluation, and discusses 'test streaming data' and 'testing set' extensively. While it mentions 'Adequate hyper-paramter analysis, provided in the supplementary', it does not explicitly describe a distinct validation dataset split within the main text for their experiments. |
| Hardware Specification | Yes | All of our experiments can be performed on a single NVIDIA Ge Force RTX 3090 card. |
| Software Dependencies | No | The paper mentions using the 'Adam (Kingma and Ba 2014) optimizer' but does not specify versions for other key software components or libraries (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | For most competing methods and our TRIBE, we leverage the Adam (Kingma and Ba 2014) optimizer with the learning rate 1e-3 in CIFAR10/100-C and Image Net-C experiments. We use a batchsize of 64 for CIFAR10/100-C and 48 for Image Net-C. Other hyper-parameters of our proposed model are listed as follow: λanc = 0.5, η = 0.0005 Kc in all datasets, in CIFAR10-C H0 = 0.05, γ = 0., in CIFAR100-C H0 = 0.2, γ = 0.1 and in Image Net-C H0 = 0.4, γ = 0.5. |