Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition
Authors: Yijie Wang, Mingjian Hong, Luwen Huangfu, Sheng Huang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach demonstrates its effectiveness across established GZSL benchmarks, seamlessly integrating into mainstream generative frameworks. Extensive experiments consistently showcase that D3GZSL elevates the performance of existing generative GZSL methods, underscoring its potential to refine zero-shot learning practices.The code is available at: https://github.com/PJBQ/D3GZSL.git |
| Researcher Affiliation | Academia | 1Chongqing University 2Fowler College of Business, San Diego State University wangyj@stu.cqu.edu.cn, hmj@cqu.edu.cn, lhuangfu@sdsu.edu, huangsheng@cqu.edu.cn |
| Pseudocode | No | The paper does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | The code is available at: https://github.com/PJBQ/D3GZSL.git |
| Open Datasets | Yes | We perform experiments on four ZSL benchmark datasets that are widely used: the Animals with Attributes1&2 (AWA1 (Lampert, Nickisch, and Harmeling 2013) & AWA2 (Xian et al. 2018a)) dataset, Caltech-UCSD Birds-200-2011 (CUB (Wah et al. 2011)) dataset, and Oxford Flowers (FLO (Nilsback and Zisserman 2008)) dataset. |
| Dataset Splits | No | The paper mentions 'training set' and 'test set' but does not explicitly provide details for a validation split (e.g., specific percentages or counts for training, validation, and test sets, or a defined cross-validation strategy). |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used (e.g., GPU/CPU models, memory details) for running its experiments, only mentioning general settings like batch size. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9') needed for replication. |
| Experiment Setup | Yes | We set the embedding dimension z to 2048 on all datasets. The classifier Cs outputs logits on all classes, and the classifier Co outputs logits on seen classes. The projector H maps softmax probabilities into a two-dimensional space that encodes both ID and OOD information. The input noise dimension w in the generator is equal to that of the corresponding attributes. In batch distillation, instances of the same class within a batch serve as positive samples for each other, while those of different class are treated as negative samples. Here are some of the parameter settings when employing f-CLSWGAN as the baseline model. We set batch size of 4096 for AWA1, 256 for CUB, 512 for FLO. The number of generated samples for each unseen category is as follows: 200 for AWA1, 5 for CUB, and 30 for FLO. We empirically set the loss weights λ = 0.0001 for AWA1, CUB. We set λ = 0.001 for FLO. |