Diversity-Enhanced and Classification-Aware Prompt Learning for Few-Shot Learning via Stable Diffusion

Authors: Gaoqin Chang, Jun Shu, Xiang Yuan, Deyu Meng

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our approach shows consistently improvements in various classification datasets, with results comparable to existing prompt designing methods. We find that replacing data generation strategy of existing zero/fewshot methods with proposed method could consistently improve downstream classification performance across different network architectures, demonstrating its model-agnostic potential for few-shot learning. This makes it possible to train an efficient downstream few-shot learning model from synthetic images generated by proposed method for real problems. Code please see: https://github.com/changxiaoqin/De Cap. ... Section 4 demonstrates experimental results and the conclusion is finally made.
Researcher Affiliation Academia Gaoqin Chang EMAIL School of Mathematics and Statistics, Ministry of Education Key Lab of Intelligent Networks and Network Security, Xi an Jiaotong University Jun Shu EMAIL School of Mathematics and Statistics, Ministry of Education Key Lab of Intelligent Networks and Network Security, Xi an Jiaotong University Pengcheng Laboratory Xiang Yuan relojeffrey@gmail.com School of Mathematics and Statistics, Ministry of Education Key Lab of Intelligent Networks and Network Security, Xi an Jiaotong University Deyu Meng EMAIL School of Mathematics and Statistics, Ministry of Education Key Lab of Intelligent Networks and Network Security, Xi an Jiaotong University Pengcheng Laboratory
Pseudocode Yes Algorithm 1 Learning Algorithm of the De Cap Method ... Algorithm 2 Get_prompt Algorithm
Open Source Code Yes Code please see: https://github.com/changxiaoqin/De Cap.
Open Datasets Yes We conduct experiments on 12 datasets: CIFAR10, STL-10, Imagenette, Pets, Caltech-101, Image Net100, Euro SAT, FGVC Aircraft, Country211, DTD, UCF101, Imagenet. Datasets details are introduced in Appendix B.1.
Dataset Splits Yes Regarding the selection of few-shot datasets, we randomly selected 10 images per class to construct the few-shot datasets. ... All the results are the average over 5 times run, with random seed in 7, 21, 42, 84, 105.
Hardware Specification Yes We do training on 8 NVIDIA A800 GPUs, with pytorch 1.12.1 and Ubuntu 20.04.
Software Dependencies Yes We do training on 8 NVIDIA A800 GPUs, with pytorch 1.12.1 and Ubuntu 20.04.
Experiment Setup Yes For inner-level training of classification model, we generated 80 images for each class and trained for 20 epochs using the Adam optimizer with a learning rate from 2e 3 to 2e 5, equipped with the cosine learning rate schedule. For outer-level training, we set the hyper-parameters of the GA algorithm as follows: popsize of 80, maxiter of 80. ... We generated 800 images for each class and fine-tune CLIP for 30 epochs. We use the the Adam W optimizer equipped with the cosine learning schedule.