Skip to content

Latest commit

 

History

History
70 lines (51 loc) · 3.55 KB

README.md

File metadata and controls

70 lines (51 loc) · 3.55 KB
Raw

DOZSL

Code and Data for the paper: "Disentangled Ontology Embedding for Zero-shot Learning". Yuxia Geng, Jiaoyan Chen, Wen Zhang, Yajing Xu, Zhuo Chen, Jeff Z. Pan, Yufeng Huang, Feiyu Xiong and Huajun Chen. KDD 2022 Research Track.

In this work, we focus on ontologies for augmenting ZSL, and propose to learn disentangled ontology embeddings to capture and utilize more fine-grained class relationships in different aspects. We also contribute a new ZSL framework named DOZSL, which contains two new ZSL solutions based on generative models and graph propagation models respectively, for effectively utilizing the disentangled ontology embeddings for zero-shot image classification (ZS-IMGC) and zero-shot KG completion (ZS-KGC) with unseen relations.

Requirements

  • python 3.5
  • PyTorch >= 1.5.0

Dataset Preparation

AwA2

Download public pre-trained image features and dataset split for AwA2, uncompress it and put the files in AWA2 folder to our folder ZS_IMGC/data/AwA2/.

ImageNet (ImNet-A, ImNet-O)

Download pre-trained image features of ImageNet classes and their class splits from here and put them to the folder ZS_IMGC/data/ImageNet/.

NELL-ZS & Wiki-ZS

Download from here and put them to the corresponding data folder.

Basic Training and Testing

Disentangled Ontology Encoder

The first thing you need to do is to train the disentangled ontology encoder, using the codes in the folders OntoEncoder/DOZSL_RD (for RD variants) and OntoEncoder/DOZSL_AGG (for AGG variants).

Steps:

  1. Running run.py in each method folder to obtain the disentangled concept embeddings;
  2. Selecting target class or relation embeddings from the trained concept embeddings by running python out_imgc.py for ZS-IMGC task and python out_kgc.py for ZS-KGC task.

Entangled ZSL Learner

With the selected class embedding or relation embedding, you can take it to perform downstream ZSL tasks using the generative model or graph propagation model.

The codes for generative model are in folder ZS_IMGC/models/DOZSL_GAN and ZS_KGC/models/DOZSL_GAN for ZS-IMGC and ZS-KGC tasks, respectively, for propagation model are in folder ZS_IMGC/models/DOZSL_GCN and ZS_KGC/models/DOZSL_GCN.

Note: you can skip the step of training ontology encoder if you just want to use the ontology embedding we learned, the embedding files have already been attached in the corresponding directories.

Baselines

  • The baselines for different ZSL methods are in the folders ZS_IMGC/models and ZS_KGC/models for ZS-IMGC and ZS-KGC tasks, respectively.
  • The baselines for different ontology embedding methods are in the folder OntoEncoder.

How to Cite

If you find this code useful, please consider citing the following paper.

@inproceedings{geng2022dozsl,
  author    = {Yuxia Geng and
               Jiaoyan Chen and
               Wen Zhang and
               Yajing Xu and
               Zhuo Chen and
               Jeff Z. Pan and
               Yufeng Huang and
               Feiyu Xiong and
               Huajun Chen},
  title     = {Disentangled Ontology Embedding for Zero-shot Learning},
  booktitle = {{KDD} '22: 28th ACM SIGKDD
Conference on Knowledge Discovery and Data Mining / Washington DC, USA, August 14-18, 2022},
  publisher = {{ACM} / {IW3C2}},
  year      = {2022}
}