Data-Free Generalized Zero-Shot Learning
Authors: Bowen Tang, Jing Zhang, Long Yan, Qian Yu, Lu Sheng, Dong Xu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our framework has been evaluated on five commonly used benchmarks for generalized ZSL, as well as 11 benchmarks for the base-to-new ZSL. The results demonstrate the superiority and effectiveness of our approach. |
| Researcher Affiliation | Academia | 1 School of Software, Beihang University 2 Department of Computer Science, The University of Hong Kong |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available in https://github.com/ylong4/DFZSL. |
| Open Datasets | Yes | Our framework is evaluated on five datasets: Attribute Pascal and Yahoo (APY) (Farhadi et al. 2009), Caltech UCSD-Birds (CUB) (Welinder et al. 2010), Oxford Flowers (FLO) (Nilsback and Zisserman 2008), SUN Attribute (SUN) (Patterson and Hays 2012), and Animals with Attributes2 (AWA2) (Xian et al. 2018a)... This includes a large-scale visual dataset, Image Net (Deng et al. 2009); a generic-objects datasets, Caltech101 (Fei-Fei 2004); five fine-grained image recognition datasets, Oxford Pets (Parkhi et al. 2012), Stanford Cars (Krause et al. 2013), Flowers102 (Nilsback and Zisserman 2008), Food101 (Bossard, Guillaumin, and Van Gool 2014) and FGVCAircraft (Maji et al. 2013); a satellite-view topographic image dataset Euro SAT (Helber et al. 2019); an action recognition dataset UCF101 (Soomro, Zamir, and Shah 2012); a texture dataset DTD (Cimpoi et al. 2014); and a scene recognition dataset SUN397 (Xiao et al. 2010). |
| Dataset Splits | Yes | We follow the splits and evaluation protocols proposed in (Xian et al. 2018a), train on base classes and then evaluated on test set that mixes the base classes and the new classes. We follow Co Co Op (Zhou et al. 2022a) to makes a half-and-half split on 11 datasets, which turns out to divide them into two non-overlapping subsets: the base classes and the new classes. |
| Hardware Specification | Yes | All experiments are performed on an NVIDIA Ge Force RTX3090, except for the Image Net, which is performed on an NVIDIA A100. |
| Software Dependencies | No | The paper mentions using specific models like CLIP and optimizers like Adam, but does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | In the black-box scenario, the class prototypes are initialized by text features, and λ is set to 1. We apply the Adam optimizer and the learning rate is set to 0.0003. In the second prompt tuning stage, we implement the light mapping network with a single-hidden-layer MLP activated by GELU... where α is the trade-off parameter that controls how much we add the shift term... and τ is the temperature used in CLIP which equals to 0.01. |