Image-free Zero-Shot Learning

This is the official repository for Image-free Classifier Injection for Zero-Shot Classification, published at ICCV 2023.

Prerequisites

1. Extracted fine-tuned CUB, AWA2, and SUN features are available here (features are originally from this paper).

2. Pre-trained ResNet101 features for CUB, AWA2, and SUN datasets are available here. We extract the features using the current version (2) of ResNet101 available via torchvision. See /utility/feature_extraction/feature_extract.py. We use the proposed datasplits ("Proposed Split Version 2.0"). Image datasets are avaiable here for CUB, AWA2, and SUN.

3. Modify default "dataroot" and "rootpath" in main.py (to point to your data path and where to save outputs, respectively).

Usage

Scripts are supplied (in the scripts folder) for experiments behind the various tables. By default, these scripts will run a single random seed and include training/loading the base classiciation model and evaluation of the specified I-(G)ZSL method. Seed number can be increased either with the --numSeeds arg, or by supplying fixed seeds with the --manualSeed arg (the latter will also speed up experiments, as base classification models can be reused).

A .yml file is supplied with the environment. (Replace version of cudatoolkit with the appropriate CUDA version of your system. Our experiments were done using CUDA/11.1 and CUDNN/8.1).

Classnames for ImageNet1K are available in /utility/.

Example usage apart from scripts include:

$ python main.py --cuda --manualSeed 0 --dataset=CUB --image_embedding=res101_finetuned --class_embedding=att --cos_sim_loss --include_unseen --num_layers 2 --beta1 0.9 --lr 0.00001 --batch_size 16 --embed_dim 2048 --strict_eval --early_stopping_slope

Citation

If you use this code, please cite

@InProceedings{Christensen_2023_ICCV,
    author    = {Christensen, Anders and Mancini, Massimiliano and Koepke, A. Sophia and Winther, Ole and Akata, Zeynep},
    title     = {Image-Free Classifier Injection for Zero-Shot Classification},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {19072-19081}
}

Note: The wDAE-GNN, Sub.Reg. and wAVG/SMO implementations are imported from their respective repositories and adapted to our usecase. If you find those parts useful, please consider citing them.

Contact

Code will receive minor updates. Questions etc. can be sent by email to

Anders Christensen

andchri@dtu.dk

Technical University of Denmark & University of Tübingen