maciejzamorski/3d-aae

Adversarial Autoencoders for Compact Representations of 3D Point Clouds

Adversarial Autoencoders for Compact Representations of 3D Point Clouds

Authors: Maciej Zamorski, Maciej Zięba, Piotr Klukowski, Rafał Nowak, Karol Kurach, Wojciech Stokowiec, and Tomasz Trzciński

Introduction

This is a PyTorch implementation for a family of 3dAAE models, a novel framework for learning continuous and binary representations of 3d point clouds based on Adversarial Autoencoder model, as presented in:

M. Zamorski, M. Zięba, et al., Adversarial Autoencoders for Compact Representations of 3D Point Clouds, arXiv preprint (2018)

Citation

@article{zamorski2018adversarial,
  title={Adversarial Autoencoders for Compact Representations of 3D Point Clouds},
  author={Zamorski, Maciej and Zi{\k{e}}ba, Maciej and Klukowski, Piotr and Nowak, Rafa{\l} and Kurach, Karol and Stokowiec, Wojciech and Trzci{\'n}ski, Tomasz},
  journal={arXiv preprint arXiv:1811.07605},
  year={2018}
}

Requirements

Stored in requirements.txt, Python dependencies are:

h5py
matplotlib
numpy
pandas
git+https://github.com/szagoruyko/pyinn.git@master
torch==0.4.1

Usage

Training

Run an experiment with:

python3.6 experiments/train.py --config settings.json

where

train.py - one of the training scripts from the experiments directory

settings.json - JSON file with training settings and hyperparameter values created as shown in example settings/hyperparams.json

Evaluation

python3.6 evaluation/find_best_epoch_on_validation.py --config settings.json

Calculates JSD distance between sampled point clouds and the validation set and presents the best epoch.

python3.6 evaluation/generate_data_for_metrics.py --config settings.json

Produce reconstructed and generated point clouds in a form of NumPy array to be used with validation methods from "Learning Representations and Generative Models For 3D Point Clouds" repository

3d-aae's Issues

Model for VAE

Are you able to make the model for the variational autoencoder available?
I noticed that it was there in your hyperparameters.json as an example but not in the model's folder.
Thanks

AAE with reparameterization trick

Hi, thanks for publishing your code!

I have a question about your implementation of the Adversarial AutoEncoder (AAE).
In your code I see you use the reparameterization trick (https://github.com/MaciejZamorski/3d-AAE/blob/master/models/aae.py), like you would if you would train a VAE.

Is there a specific reason to do this? Because I'd think this is not necessary?

Kind regards

Rembert

Has anyone reproduce the result of 3D-AAE and get JSD~=0.014 for the chair datasets

Hello,
If so , can you provide the hyperparameter for training?

Earth mover distance

I try to transfer the emd implementation to my own project, but it occurs some wired things.

The inputs are input point cloud and reconstructed point cloud with dimension [batch size, num points, num dim(3)].

But after some iterations, the reconstructed point cloud is very comfused.

Should I pay some attention for using this end? Thanks

Regards
Xin

How to get Compact binary embeddings in your code ?