(* denotes equal contribution)

Webpage: https://akshitac8.github.io/GAN_MLZSL/

This repository contains the implementation of Generative Multi-Label Zero-Shot Learning.

In this work, we tackle the problem of synthesizing multi-label features in the context of zero-shot setting for recognition all (un)seen labels with a novel training mechanism.

The codebase is built on PyTorch 1.1.0 and tested on Ubuntu 16.04 environment (Python3.6, CUDA9.0, cuDNN7.5).

For installing, follow these intructions

conda create -n mlzsl python=3.6
conda activate mlzsl
conda install pytorch=1.1 torchvision=0.3 cudatoolkit=9.0 -c pytorch
pip install matplotlib scikit-image scikit-learn opencv-python yacs joblib natsort tqdm pandas h5py==2.10.0

Download the NUS-WIDE features, tags and other required training files from the drive link shared below.

link: https://drive.google.com/drive/folders/1tCo-xawWrnGQGaWYJEKQOQ31ts__rAse?usp=sharing

Extract them in the ./datasets folder.

Download the custom dataset in the same data folder. Please make sure to convert your custom dataset in the same format as NUS-WIDE.

python preprocess.py --image_dir data/custom_data/ --output_dir data/custom_data_jsons/ --train_json custom_data_train  --test_json custom_data_test

Above preprocessing step will create train and test jsons with ZSL and GZSL requirements.

The train and test jsons are used as an input when running feature extraction code.

python extract_4096_features.py --train_json custom_data_train  --test_json custom_data_test --gpu

Above feature extraction will save features in .h5 format which is used for training for our CLF model.

To train and evaluate zero-shot learning model on full NUS-WIDE dataset, run:

sh ./scripts/train_nus_wide.sh

We also include the checkpoint of the zero-shot generative model on NUS-WIDE for fast evaluation in weights folder. Please download the pretrained weights according to the intructions within the folder. To reproduce results, run:

sh ./scripts/eval_nus_wide.sh

If this code is helpful for your research, we would appreciate if you cite the work:

@article{gupta2021generative,
  title={Generative Multi-Label Zero-Shot Learning},
  author={Gupta, Akshita and Narayan, Sanath and Khan, Salman and Khan, Fahad Shahbaz and Shao, Ling and van de Weijer, Joost},
  journal={arXiv preprint arXiv:2101.11606},
  year={2021}

I thank Dat Huynh for discussions and feedback regarding the evaluation protocol and sharing details for the baseline zero-shot methods. I thank Aditya Arora for suggestions on the figure aesthetics.