PyTorch implementation of slot-attention-based architectures for object discover. This code uses Sacred to define and run experiments.
Two models are provided, the standard Slot-Attention model (Locatello et al, 2020) and SLATE (Singh et al, 2022). Additionally, users can use implicit gradients as proposed in (Chang et al, 2022).
To replicate the experiments in the above articles, users should run the script bin/replicate
inside the experiments
folder. Only 3DShapes
and Tetrominoes
datasets are provided, with raw data found at their respective repos.
The environment can be installed using the following command:
conda env create -f torchlab-env.yml
The relevant .yaml
config file can be found here
We thank the authors of both the Slot-Attention and Slate models for making their code available, wich allowed us to reproduce the results.
@misc{locatello2020objectcentric,
title = {Object-Centric Learning with Slot Attention},
author = {Francesco Locatello and Dirk Weissenborn and Thomas Unterthiner and Aravindh Mahendran and Georg Heigold and Jakob Uszkoreit and Alexey Dosovitskiy and Thomas Kipf},
year = {2020},
eprint = {2006.15055},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}
@inproceedings{
singh2022illiterate,
title={Illiterate DALL-E Learns to Compose},
author={Gautam Singh and Fei Deng and Sungjin Ahn},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=h0OYV0We3oh}
}