HSVA-NeurIPS-21

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

Preparation

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.

Training

To Train the HSVA in the GZSL or CZSL setting, please run the commands in the following:

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

Testing

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

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

Datasets	U	S	H	acc
AWA1	61.1	75.2	67.4	70.6
AWA2	57.8	79.3	66.9	--
CUB	51.9	59.5	55.5	62.8
SUN	48.6	39.0	43.3	63.8

Citation

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}    
}

Ackowledgement

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

1. CADA-VAE
2. MCD-DA