Skip to content

[IJCNN 2024] Masked Multi-Query Slot Attention for Unsupervised Object Discovery, 2024 International Joint Conference on Neural Networks

License

Notifications You must be signed in to change notification settings

rishavpramanik/maskedmultiqueryslot

Repository files navigation

Masked Multi-Query Slot Attention

"Masked Multi-Query Slot Attention for Unsupervised Object Discovery" - accepted for oral presentation in 2024 International Joint Conference on Neural Networks Yokohama, Japan.

Access the paper: Arxiv IEEE Xplore

Datasets Used:

PASCAL VOC 2012 Click here

Requirements

  • Python >= 3.8
  • PyTorch >= 1.7.1
  • Pytorch Lightning >= 1.1.4
  • CUDA enabled computing device
  • For more requirements please consult requirements.txt

Instructions to run the code:

  1. Download the repository and install the required packages:
pip3 install -r requirements.txt
  1. Unzip the data in a folder of your choice
tar -xf yourdirectory/VOCtrainval_11-May-2012.tar -C $SLURM_TMPDIR/yourdirectory
  1. The train2 file is sufficent to run the code
torchrun --nproc_per_node=4 --nnodes=1 scripts/train2.py

Edit the parameters before you start in params.py and other required places before you start

Citation

@INPROCEEDINGS{pramanik2024masked,
AUTHOR="Rishav Pramanik and Jos{\'e}-Fabian {Villa-V{\'a}squez} and Marco Pedersoli",
TITLE="Masked {Multi-Query} Slot Attention for Unsupervised Object Discovery",
BOOKTITLE="2024 International Joint Conference on Neural Networks (IJCNN) (IJCNN 2024)",
ADDRESS="Yokohama, Japan",
DAYS=28,
MONTH=jun,
YEAR=2024,
}

Acknowledgements

We greatly thank the authors of https://github.com/amazon-science/object-centric-learning-framework/tree/main and https://github.com/imbue-ai/slot_attention/tree/master for their code which had helped us in our work

About

[IJCNN 2024] Masked Multi-Query Slot Attention for Unsupervised Object Discovery, 2024 International Joint Conference on Neural Networks

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages