Generalized Zero-Shot Learning via Disentangled Representation
Authors: Xiangyu Li, Zhe Xu, Kun Wei, Cheng Deng1966-1974
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method outperforms the state-of-the-art approaches on four challenging benchmark datasets. In this section, all datasets and evaluation protocol are introduced in detail. In addition, we present the the implementation details as well as the comparison of experimental results with other state-of-the-art methods. Eventually, ablation study proves the effectiveness our method. |
| Researcher Affiliation | Academia | Xiangyu Li*, Zhe Xu*, Kun Wei, Cheng Deng School of Electronic Engineering, Xidian University, Xi an 710071, China {xdu xy Li, zhexu}@stu.xidian.edu.cn, {weikunsk, chdeng.xd}@gmail.com |
| Pseudocode | No | The paper describes the methodology in prose and with diagrams, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing open-source code or provide a link to a code repository for the described methodology. |
| Open Datasets | Yes | We evaluate our model on four popular datasets: Caltech UCSD-Birds 200-2011 dataset (CUB) (Welinder et al. 2010), Animals with Attributes 1 (AWA1) (Lampert, Nickisch, and Harmeling 2009) and 2 (AWA2) (Xian et al. 2018a), SUN Attribute dataset (SUN) (Patterson and Hays 2012). |
| Dataset Splits | Yes | To avoid violating the zero-shot setting, we adopt the typical training splits proposed by (Xian et al. 2018a) for training split so that test samples can be disjoint from training samples which Res Net-101 is trained with on each dataset. Due to the difference between each dataset, the proportion of seen classes samples and unseen classes samples is different and inspired the results of many experiments we have made, we use a fixed dataset with 200 samples per seen class and 400 samples per unseen class in CUB dataset, 200, 460 in AWA1 dataset, 200, 480 in AWA2 dataset, and 200, 410 in SUN dataset. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions 'Py Torch(Paszke et al. 2019)' and 'ADAM optimizer (Kingma and Ba 2014)' but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We utilize 1560, 1660 hidden units for encoder Ev and 1450, 660 hidden units for encoder Es. The size of latent feature is implemented as 64 in the whole datasets. Our approach is implemented with Py Torch(Paszke et al. 2019) and optimized by ADAM optimizer (Kingma and Ba 2014). In addition, we set learning rate as 0.00015, batch size as 50 and epochs as 150. |