Canonical Capsules: Unsupervised Capsules in Canonical Pose

Introduction

⚠️ the source code is subject to changes up to June 1st, 2021 (official release date)

This is the official repository for the PyTorch implementation of "Canonical Capsules: Unsupervised Capsules in Canonical Pose" by Weiwei Sun*, Andrea Tagliasacchi*, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, Kwang Moo Yi.

Download links

• Project Website
• PDF (arXiv)
• PDF (github copy)

Citation

⚠️ If you use this source core or data in your research (in any shape or format), we require you to cite our paper as:

@conference{sun2020canonical,
   title={Canonical Capsules: Unsupervised Capsules in Canonical Pose},
   author={Weiwei Sun and Andrea Tagliasacchi and Boyang Deng and 
           Sara Sabour and Soroosh Yazdani and Geoffrey Hinton and
           Kwang Moo Yi},
   booktitle={arXiv preprint},
   publisher_page={https://arxiv.org/abs/2012.04718},
   year={2020}
}

Requirements

Please install dependencies with the provided environment.yml:

conda env create -f environment.yml

Datasets

• We use the ShapeNet dataset as in AtlasNetV2: download the data from AtlasNetV2's official repo and convert the downloaded data into h5 files with the provided script (i.e., data_utils/ShapeNetLoader.py).

• For faster experimentation, please use our 2D planes dataset, which we generated from ShapeNet (please cite both our paper, as well as ShapeNet if you use this dataset).

Training/testing (2D)

To train the model on 2D planes (training of network takes only 50 epochs, and one epoch takes approximately 2.5 minutes on an NVIDIA GTX 1080 Ti):

./main.py --log_dir=plane_dim2 --indim=2 --scheduler=5

To visualize the decompostion and reconstruction:

./main.py --save_dir=gifs_plane2d --indim=2 --scheduler=5 --mode=vis --pt_file=logs/plane_dim2/checkpoint.pth

Training/testing (3D)

To train the model on the 3D dataset:

./main.py --log_dir=plane_dim3 --indim=3 --cat_id=-1

We test the model with:

./main.py --log_dir=plane_dim3 --indim=3 --cat_id=-1 --mode=test

Note that the option cat_id indicates the category id to be used to load the corresponding h5 files (this look-up table):

Pre-trained models (3D)

We release the 3D pretrained models for both single categy (airplanes), as well as multi-category (all 13 classes).

Classification

We plan to release our code for the classification experiments in the paper as well (ETA: by the end of Janurary).