Skip to content

Latest commit

 

History

History
76 lines (43 loc) · 2.59 KB

README.md

File metadata and controls

76 lines (43 loc) · 2.59 KB

SCALOR

This repository is the official implementation of "SCALOR: Generative World Models with Scalable Object Representations" by Jindong Jiang*, Sepehr Janghorbani*, Gerard de Melo, and Sungjin Ahn; accepted at the International Conference on Learning Representations (ICLR) 2020. Project Website

Architecture

Requirements

To install requirements:

conda env create -f environment.yml

To activate:

conda activate scalor_env

Dataset

The "Grand Central Station" dataset can be downloaded here. After downloading the file, extract the dataset using this command:

tar -xf grandcentralframes.tar.gz -C /path/to/dataset/

Training

To train SCALOR with default settings, run this command:

python train.py --data-dir /path/to/dataset/

Results

toy

natural-scene

Using SCALOR in your project

Foreground not working

If you find the background module explains everything in the image and the foreground module is turned off, first check the following two settings:

  1. The num_cell_h and num_cell_w in common.py. If the objects in the scene are densely positioned in a local area, the number of cells (num_cell_h and num_cell_w) should be larger to provide enough cells in that local area.

  2. The max_num_obj in common.py. In the early training stage, this number is higher the better (smaller than the total number of cells) since it allows more activated cells to accelerate the foreground training. Feel free to reduce it later.

Additionally, I also added the following two settings in the code. Feel free to try any of them:

  1. Using a weaker background decoder, one option is to set the using_bg_sbd flag to True in common.py.

  2. Using a training curriculum in the early training stage. This can be done by setting the phase_bg_alpha_curriculum to True in common.py.

Feel free to let me know if you face any other problems when adopting SCALOR in your project.

Citation

@inproceedings{JiangJanghorbaniDeMeloAhn2020SCALOR,
  title={SCALOR: Generative World Models with Scalable Object Representations},
  author={Jiang, Jindong and Janghorbani, Sepehr and De Melo, Gerard and Ahn, Sungjin},
  booktitle={International Conference on Learning Representations},
  year={2019}
}