Meta-Adapter: An Online Few-shot Learner for Vision-Language Model

Authors: cheng cheng, Lin Song, Ruoyi Xue, Hang Wang, Hongbin Sun, Yixiao Ge, Ying Shan

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

Reproducibility Variable Result LLM Response
Research Type Experimental The extensive experiments demonstrate the effectiveness and efficiency of the Meta-Adapter on image classification, object detection, and segmentation. To verify the generalizability of Meta-Adapter, we conduct a series of ablation studies, including cross-category generalization within a certain dataset, cross-dataset generalization, and cross-task generalization which explores the potential of Meta-Adapter in downstream tasks.
Researcher Affiliation Collaboration 1 Xi an Jiao Tong University 2Tencent AI Lab
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code, nor does it include an explicit code release statement or repository link.
Open Datasets Yes For cross-category generalization experiments, we use 8 representative image classification datasets: Image Net [29], FGVCAircraft [51], Oxford Pets [52], SUN397 [53], UCF101 [54], Caltech101 [55], DTD [56], and Euro SAT [57]... As for the cross-dataset generalization experiment, Image Net is further utilized as the source dataset and its three variants are treated as target datasets, i.e., Image Net-A [58], Image Net-R [59], and Image Net-Sketch [60]. Moreover, to explore the potential of Meta-Adapter on open-vocabulary detection, we conduct experiments on LVIS [61].
Dataset Splits Yes Specifically, for cross-category generalization , we split the full categories of each dataset into base and novel sets according to the per-category accuracy predicted by Zero-shot CLIP, that is, the base set contains easy samples and the novel set contains hard samples. This dataset split strategy simulates a rather difficult situation to verify whether Meta-Adapter is able to learn the dataset-irrelevant approach, especially for hard samples.
Hardware Specification Yes The time is measured on a Tesla V100 GPU.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) needed to replicate the experiment.
Experiment Setup Yes We optimize the Meta-Adapter on the base set with a batch size of 64 and use Adam W optimizer [62] with a learning rate of 0.0001 and a cosine scheduler for 5 epochs.