Paper: Refinement Matters: Textual Description Needs to be Refined for Zero-shot Learning (EMNLP-2022 (Findings))

Zero-Shot Learning

• Download the AWA2 dataset using link1/link2
• Provide the path for the AWA2 dataset in the 'DATA_DIR' variable in the notebook
• Run the CARNet.ipynb notebook to get the ZSL result of the proposed CARNet method for the AWA2 dataset
• Variables 'USE_ATTRIBUTE_REFINEMENT' and 'USE_CIRCLE_LOSS' can be set to 'False' to get results without attribute refinement and circle loss

Continual Zero-Shot Learning

• Download the SUN dataset using link1/link2
• Provide the path for the SUN dataset in the 'DATA_DIR' variable in the notebook
• Run the CARNet_Continual_ZSL.ipynb notebook to get the Continual ZSL result of the proposed CARNet method for the SUN dataset

Note: Due to some unknown reasons, the code in the notebook might be repeated twice when viewed from the browser directly from the link. However, if you download the notebook and view it in the jupyter notebook this problem does not occur