This is the implementation of “HSVA: Hierarchical Semantic-Visual Adaptation for Zero-Shot Learning” in Pytorch. The paper will appear in NeurIPS 2021.

1. Datasets can be downloaded Here. Put it in the data dir.
2. Download our pretrained models from google drive Here, including CUB, SUN, AWA1 and AWA2 models. Put it in the result dir. Note that we just provide one pre-trained model for every dataset.

To test the results for GZSL or CZSL, please run the commands in the following:

CUDA_VISIBLE_DEVICES="2" python test.py --dataset CUB --latent_size 64 --generalized True
CUDA_VISIBLE_DEVICES="2" python test.py --dataset SUN --latent_size 128 --generalized True
CUDA_VISIBLE_DEVICES="2" python test.py --dataset AWA1 --latent_size 64 --generalized True
CUDA_VISIBLE_DEVICES="2" python test.py --dataset AWA2 --latent_size 64 --generalized True

--gdataset test dataset, i.e., CUB, SUN, AWA1, and AWA2.

--generalized test for GZSL (True) or CZSL (False).

Results of our released model using various evaluation protocols on four datasets, both in conventional ZSL (CZSL) and generalized ZSL (GZSL) setting.

If this work is helpful for you, please cite our paper.

@inproceedings{Chen2021HSVA,  
  title={HSVA: Hierarchical Semantic-Visual Adaptation for Zero-Shot Learning},    
  author={Chen, Shiming and Xie, Guo-Sen and Peng, Qinmu and Liu, Yang and Sun, Baigui and Li, Hao and You, Xinge and Shao,  Ling},    
  booktitle={35th Conference on Neural Information Processing Systems (NeurIPS)}, 
  year={2021}    
}

We thank the following repos providing helpful components in our work.

1. CADA-VAE
2. MCD-DA