Candidate Pseudolabel Learning: Enhancing Vision-Language Models by Prompt Tuning with Unlabeled Data

Authors: Jiahan Zhang, Qi Wei, Feng Liu, Lei Feng

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on nine benchmark datasets with three learning paradigms demonstrate the effectiveness of our method.
Researcher Affiliation Academia 1 Singapore University of Technology and Design, Singapore 2 Nanyang Technological University, Singapore 3 University of Melbourne, Australia.
Pseudocode Yes Algorithm 1 Top-K Selection Process in Each Iteration
Open Source Code Yes Our code can be found here.
Open Datasets Yes We conduct an extensive evaluation of our method on nine classification datasets from diverse domains, including FGVC-Aircraft (Maji et al., 2013), Euro SAT (Helber et al., 2019), CUB (Wah et al., 2011), Flowers102 (Nilsback & Zisserman, 2008), RESISC45 (Cheng et al., 2017), DTD (Cimpoi et al., 2014), CALTECH-101 (Fei-Fei et al., 2004), UCF-101 (Soomro et al., 2012), and CIFAR-100 (Krizhevsky et al., 2009).
Dataset Splits No The paper provides 'Training set size' and 'Testing set size' in Table 8 for various datasets. For Semi-Supervised Learning, it mentions using 'two labeled samples per class' but does not explicitly define a separate 'validation set' with specific percentages or counts for hyperparameter tuning.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or cloud instance types used for running the experiments.
Software Dependencies No The paper mentions 'Optimizer SGD' and 'Network Vi T-B / 32' but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes Table 8: Detailed settings for experiments in Section 4. Training procedure: Network Vi T-B / 32, Batch size 64, Epoch 50 where first two epochs are set for warmup, Optimizer SGD, Momentum 0.9, Learning rate (LR) 0.02, Weight decay 5e-2, LR scheduler Cosine Annealing LR. Hyperparameters: α in intra-instance label selection (e.g., 0.60 for Flowers102), β in inter-instance label selection (e.g., 0.99 for Flowers102).