TransHP: Image Classification with Hierarchical Prompting

Authors: Wenhao Wang, Yifan Sun, Wei Li, Yi Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that Trans HP improves image classification on accuracy (e.g., improving Vi T-B/16 by +2.83% Image Net classification accuracy), training data efficiency (e.g., +12.69% improvement under 10% Image Net training data), and model explainability. Moreover, Trans HP also performs favorably against prior HIC methods, showing that Trans HP well exploits the hierarchical information.
Researcher Affiliation Collaboration Wenhao Wang Re LER, University of Technology Sydney Yifan Sun Baidu Inc. Wei Li Zhejiang University Yi Yang Zhejiang University
Pseudocode No The paper describes the model architecture and steps but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available at: https://github.com/Wang Wenhao0716/Trans HP.
Open Datasets Yes We evaluate the proposed Trans HP on five datasets with hierarchical labels, i.e., Image Net [10], i Naturalist-2018 [11], i Naturalist-2019 [11], CIFAR-100 [25], and Deep Fashion-inshop [26].
Dataset Splits Yes To this end, we randomly select 1/10, 1/5, and 1/2 training data from each class in Image Net (while keeping the validation set untouched).
Hardware Specification Yes We train it for 300 epochs on 8 Nvidia A100 GPUs and Py Torch.
Software Dependencies No The paper mentions 'Py Torch' as a software dependency but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes The base learning rate is 0.001 with cosine learning rate. We set the batch size, the weight decay and the number of warming up epochs as 1,024, 0.05 and 5, respectively.