Bi-directional Distribution Alignment for Transductive Zero-Shot Learning

Zero-shot learning (ZSL) suffers intensely from the domain shift issue, i.e., the mismatch (or misalignment) between the true and learned data distributions for classes without training data (unseen classes). By learning additionally from unlabelled data collected for the unseen classes, transductive ZSL (TZSL) could reduce the shift but only to a certain extent. To improve TZSL, we propose a novel approach Bi-VAEGAN which strengthens the distribution alignment between the visual space and an auxiliary space.

Requirements

The code is built upon library of TF-VAEGAN, and versions newer than TF-VAEGAN's prerequisites should work fine with this code.

numpy
torch
sklearn
tensorboardX 

Data preparation

Download extracted AwA1, AwA2, CUB, FLO, SUN feature datasets from here, these data are mostly the same with TF-VAEGAN:

https://drive.google.com/file/d/1Uax5IbQCaZhh7BAbwxos7Gq8SAKKbWbe/view?usp=share_link

Training with known class prior

Check the scripts in scripts directory,

bash scripts/awa.sh

Training with unknown class prior

Specify --unknown_classDistribution in the script and choose a prior estimation method like --prior_estimation 'CPE'.

#bash scripts/awa.sh
# --prior_estimation prior estimation strategy. Should be within 'CPE', 'classifier' or 'BBSE'
# --ind_epoch number of inductive training epochs before switching to transductive training
    ...
    --unknown_classDistribution \
    --ind_epoch 3 \ 
    --prior_estimation 'CPE'
    ...

Result

TZSL accuracy under known class prior:

Cite

@inproceedings{Bi-VAEGAN, title={Bi-directional Distribution Alignment for Transductive Zero-Shot Learning}, author={Wang, Zhicai and Hao, Yanbin and Mu, Tingting and Li, Ouxiang and Wang, Shuo and He, Xiangnan}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={19893--19902}, year={2023} }