Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection
Authors: Fanhu Zeng, Zhen Cheng, Fei Zhu, Hongxin Wei, Xu-yao Zhang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments demonstrate the effectiveness and potential of our method. Notably, our method reduces average FPR95 by 5.17% against state-of-the-art method in 4-shot tuning on challenging Image Net1k dataset, even outperforming 16-shot results of previous methods. Code is released at https://github.com/Aurora Zengfh/Local-Prompt. |
| Researcher Affiliation | Academia | 1State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences 2School of Artificial Intelligence, University of Chinese Academy of Sciences 3Centre for Artificial Intelligence and Robotics, HKISI-CAS 4Department of Statistics and Data Science, Southern University of Science and Technology EMAIL EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical formulations and textual explanations in Section 4 'METHODOLOGY', but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is released at https://github.com/Aurora Zengfh/Local-Prompt. |
| Open Datasets | Yes | Datasets. Following existing works (Ming et al., 2022; Miyai et al., 2023b;a), large scale Image Net1K (Deng et al., 2009) along with a 100-category subset of it denoted as Image Net-100 are used as ID dataset and OOD dataset is a combination of i Naturalist (Van Horn et al., 2018), SUN (Xiao et al., 2010), Places (Zhou et al., 2017) and Texture (Cimpoi et al., 2014). |
| Dataset Splits | Yes | Under few-shot setting, a small proportion of images in each class (e.g., 1, 4, or 16 images) are extracted for training. ... For evaluation, we use validation set of Image Net-1k, which is composed of 50000 images from 1000 categories. |
| Hardware Specification | Yes | We use one single NVIDIA A6000 to run all experiments. |
| Software Dependencies | No | The paper mentions using "CLIP-Base/16 as the backbone" and an "SGD optimizer" but does not specify software library versions (e.g., PyTorch version, specific dependencies with version numbers). |
| Experiment Setup | Yes | Number of learnable context words is 16. We use SGD optimizer to train the model with 30 epochs. Learning rate is 2 10 3 with a cosine schedule and batch size is 256. m is set to 24 by default. We empirically set positive/negative augmentation to be 8/1, which are discussed thoroughly below. Training and testing k are 50 and 10, respectively in main results. |