Official PyTorch Implementation of Adaptive Slot Attention: Object Discovery with Dynamic Slot Number

Adaptive Slot Attention: Object Discovery with Dynamic Slot Number
Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu†, Francesco Locatello, Zheng Zhang

This is the official implementation of the CVPR'24 paper [Adaptive Slot Attention: Object Discovery with Dynamic Slot Number](CVPR 2024 Open Access Repository (thecvf.com)).

Introduction

Object-centric learning (OCL) uses slots to extract object representations, enhancing flexibility and interpretability. Slot attention, a common OCL method, refines slot representations with attention mechanisms but requires predefined slot numbers, ignoring object variability. To address this, a novel complexity-aware object auto-encoder framework introduces adaptive slot attention (AdaSlot), dynamically determining the optimal slot count based on data content through a discrete slot sampling module. A masked slot decoder suppresses unselected slots during decoding. Extensive testing shows this framework matches or exceeds fixed-slot models, adapting slot numbers based on instance complexity and promising further research opportunities.

Development Setup

Installing AdaSlot requires at least python3.8. Installation can be done using poetry. After installing poetry, check out the repo and setup a development environment:

# install python3.8
sudo apt update
sudo apt install software-properties-common
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt install python3.8

# install poetry with python3.8
curl -sSL https://install.python-poetry.org | python3.8 - --version 1.2.0
## add poetry to environment variable

# create virtual environment with poetry
cd $code_path
poetry install -E timm

This installs the ocl package and the cli scripts used for running experiments in a poetry managed virtual environment. Activate the poetry virtual environment poetry shell before running the experiments.

Running experiments

Experiments are defined in the folder configs/experiment and can be run by setting the experiment variable. For example, if we run OC-MOT on Cater dataset, we can follow:

poetry shell

python -m ocl.cli.train +experiment=projects/bridging/dinosaur/movi_e_feat_rec_vitb16.yaml
python -m ocl.cli.train +experiment=projects/bridging/dinosaur/movi_e_feat_rec_vitb16_adaslot.yaml +load_model_weight=PATH-TO-KMAX-SLOT-CHECKPOINT
python -m ocl.cli.eval +experiment=projects/bridging/dinosaur/movi_e_feat_rec_vitb16_adaslot_eval.yaml ++load_checkpoint=PATH-TO-ADASLOT-CHECKPOINT

python -m ocl.cli.train +experiment=projects/bridging/dinosaur/movi_c_feat_rec_vitb16.yaml
python -m ocl.cli.train +experiment=projects/bridging/dinosaur/movi_c_feat_rec_vitb16_adaslot.yaml +load_model_weight=PATH-TO-KMAX-SLOT-CHECKPOINT
python -m ocl.cli.eval +experiment=projects/bridging/dinosaur/movi_c_feat_rec_vitb16_adaslot_eval.yaml ++load_checkpoint=PATH-TO-ADASLOT-CHECKPOINT

python -m ocl.cli.train +experiment=projects/bridging/dinosaur/coco_feat_rec_dino_base16.yaml
python -m ocl.cli.train +experiment=projects/bridging/dinosaur/coco_feat_rec_dino_base16_adaslot.yaml +load_model_weight=PATH-TO-KMAX-SLOT-CHECKPOINT
python -m ocl.cli.eval +experiment=projects/bridging/dinosaur/coco_feat_rec_dino_base16_adaslot_eval.yaml ++load_checkpoint=PATH-TO-ADASLOT-CHECKPOINT

python -m ocl.cli.train +experiment=slot_attention/clevr10.yaml
python -m ocl.cli.train +experiment=slot_attention/clevr10_adaslot.yaml +load_model_weight=PATH-TO-KMAX-SLOT-CHECKPOINT
python -m ocl.cli.eval +experiment=slot_attention/clevr10_adaslot_eval.yaml ++load_checkpoint=PATH-TO-ADASLOT-CHECKPOINT

The result is saved in a timestamped subdirectory in outputs/<experiment_name>, i.e. outputs/OC-MOT/cater/<date>_<time> in the above case. The prefix path outputs can be configured using the experiment.root_output_path variable.

Citation

Please cite our paper if you find this repo useful!

@inproceedings{fan2024adaptive,
  title={Adaptive slot attention: Object discovery with dynamic slot number},
  author={Fan, Ke and Bai, Zechen and Xiao, Tianjun and He, Tong and Horn, Max and Fu, Yanwei and Locatello, Francesco and Zhang, Zheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={23062--23071},
  year={2024}
}

Related projects that this paper is developed upon:

@misc{oclf,
  author = {Max Horn and Maximilian Seitzer and Andrii Zadaianchuk and Zixu Zhao and Dominik Zietlow and Florian Wenzel and Tianjun Xiao},
  title = {Object Centric Learning Framework (version 0.1)},
  year  = {2023},
  url   = {https://github.com/amazon-science/object-centric-learning-framework},
}