Incremental Embedding Learning via Zero-Shot Translation
Authors: Kun Wei, Cheng Deng, Xu Yang, Maosen Li10254-10262
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on CUB-200-2011 and CIFAR100, and the experiment results prove the effectiveness of our method. |
| Researcher Affiliation | Academia | Kun Wei, Cheng Deng , Xu Yang, and Maosen Li School of Electronic Engineering, Xidian University, Xian 710071, China {weikunsk, chdeng.xd, xuyang.xd, maosenli95}@gmail.com |
| Pseudocode | No | The paper describes the training and inference steps in text and with equations but does not provide a formal pseudocode block or algorithm. |
| Open Source Code | Yes | The code of our method has been released in https://github.com/Drkun/ZSTCI. |
| Open Datasets | Yes | We evaluate the methods on two popular datasets: CUB-200-2011 (CUB) (Wah et al. 2011) and CIFAR100 (Krizhevsky, Hinton et al. 2019). |
| Dataset Splits | No | The paper states: "All these datasets are divided by classes into ten tasks randomly and the random seed is set as 1993." It also describes a "testing process" but does not specify train/validation/test splits with percentages or absolute sample counts for reproducibility within each task or overall. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used, such as GPU or CPU models, or memory specifications. |
| Software Dependencies | No | The paper mentions "All models are implemented with Pytorch" but does not specify a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | As for embedding network, Res Net-18 (He et al. 2016) is selected as the backbone network pre-trained from Image Net (Deng et al. 2009) for CUB. In addition, Res Net-32 is adopted for CIFAR100, which is without pre-training. A triplet loss is employed to regularize the learning process of embedding network. The training images are resized to 256 256 for CUB and 32 32 for CIFAR100, then randomly cropped and flipped. The epochs and learning rates are set to 50 and 1e-5 for CUB and CIFAR100 respectively. The dimension of final embeddings normalized is 512. All models are implemented with Pytorch. Adam optimizer (Kingma and Ba 2014) is employed to optimize the models and the batch size for all experiments is set to 32. As for zero-shot translation network, the mapping models are two-layer fully-connected networks, and the dimension of hidden layer is 1024. The epoch and batch size are set to 100 and 128 for CUB, 50 and 128 for CIFAR100. In addition, the learning rate is set to 0.002 and the model is optimized by Adam optimizer. γ, β and δ are set to 1000, 100 and 100 for CUB, 200, 100 and 100 for CIFAR100 respectively. |