Relation-CZSL

This repo contains the official implementation of paper "Relation-aware Compositional Zero-shot Learning for Attribute-Object Pair Recognition".

Requirements

• Ubuntu 18.04
• CUDA 10.0+
• A CUDA-compatible GPU with 8GB+ VRAM
• Anaconda/Miniconda
• 20GB HDD space

Setup

• Install Anaconda.
• Create an Anaconda environment using the configuration pytorch.yaml provided in the repo:

conda env create -n RCZL -f pytorch.yaml

• Activate the environment.

conda activate RCZL

Data and Model Preparation

• Follow the instructions in TMN to prepare the data.
• For faster training and evaluation, download precomputed features, and place the file under data/[mitstates|ut-zap50k]/. Alternatively, remove the --pre_feat flag from eval.sh if you don't want to use the these features.
• Download the weights of models here.

Evaluation

Modify variables at line 4-8 in eval.sh to a proper batch size (depending on the VRAM you have), the path to the model, and the path to the data etc. Then run

bash eval.sh

Training

Change the datapath in line3 in train_[utzap50k|mitstates].sh to the path you stored data, then run

bash train_utzap50k.sh
bash train_mitstates.sh

to start the training.

Reference

If you find this repository helpful in your research, please cite the following paper:

@article{Xu2021RZSL,
  author={Xu, Ziwei and Wang, Guangzhi and Wong, Yongkang and Kankanhalli, Mohan S.},
  journal={IEEE Transactions on Multimedia}, 
  title={Relation-aware Compositional Zero-shot Learning for Attribute-Object Pair Recognition}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMM.2021.3104411}
}