LPT: Long-tailed Prompt Tuning for Image Classification

Authors: Bowen Dong, Pan Zhou, Shuicheng YAN, Wangmeng Zuo

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that on various long-tailed benchmarks, with only 1.1% extra trainable parameters, LPT achieves comparable or higher performance than previous whole model fine-tuning methods, and is more robust to domain-shift.
Researcher Affiliation Academia Bowen Dong1 Pan Zhou2 Shuicheng Yan2 Wangmeng Zuo1,3 1Harbin Institute of Technology 2National University of Singapore 3 Peng Cheng Laboratory {cndongsky,panzhou3,shuicheng.yan}@gmail.com, wmzuo@hit.edu.cn
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is publicly available at https://github.com/DongSky/LPT.
Open Datasets Yes CIFAR100-LT is a subset of the original CIFAR-100 (Krizhevsky, 2009)... Places-LT (Liu et al., 2019) is the subset of Places-365 dataset Zhou et al. (2017a)... i Naturalist 2018 (Van Horn et al., 2018)... Image Net-Sketch (Wang et al., 2022)... Image Net-LT (Liu et al., 2019).
Dataset Splits No The paper mentions using train, val, and test sets for Places-LT and iNaturalist 2018, and generating long-tailed training data for CIFAR100-LT, but it does not specify the exact split percentages, sample counts, or cite a predefined split methodology for these datasets, which are necessary for full reproducibility of data partitioning.
Hardware Specification Yes All programs are implemented by Py Torch toolkit (Paszke et al., 2019), and all experiments are conducted on a single RTX A6000 GPU.
Software Dependencies No The paper mentions 'Py Torch toolkit (Paszke et al., 2019)' but does not provide specific version numbers for PyTorch or other software dependencies.
Experiment Setup Yes For shared prompt, we simply set the default length of prompt as 10... we use SGD optimizer with momentum of 0.9... the initial learning rate is set to be 0.002 B 256... we set the weight decay as 1e-4... For Places-LT... we optimize phase 1 and phase 2 of LPT for E = 40 epochs... For asymmetric GCL loss, we set λ+ and λ as 0 and 4, respectively. And for phase 2, we set the initialized weight γ used in {I}ins as 0.5. In all experiments, the training and testing images are resized to 224 224.