Skip to content

Latest commit

 

History

History
169 lines (129 loc) · 8.37 KB

README.md

File metadata and controls

169 lines (129 loc) · 8.37 KB

Geometric Deep Learning for Design Applications (GDL4DesignApps)

teaser

GDL4DesignApps is a software repository for training and applying geometric deep learning models in engineering design and optimization tasks. The software is an outcome of the research performed in the framework of the project Experience-based Computation: Learning to Optimise (ECOLE). ECOLE is an Innovative Training Network that is part of the Marie Skłodowska-Curie Actions (ITN-MSCA, Grant number 766186) and funded by the Horizon2020 program.

Markdown Monster icon

Software Pre-requisites

We developed our software in a machine with Ubuntu 18.04 utilizing Python 3.6.13 in an Anaconda environment with the following libraries installed:

Library Version Library Version
Numpy 1.19.1 tensorflow-gpu 1.14.0
scikit-learn 0.23.2 TFLearn 0.3.2
pandas 1.1.0 cudatoolkit 10.1.168
matplotlib 1.19.1 cudnn 7.6.5
pyvista 0.31.3 Platypus 1.0.4
seaborn 0.11.2 pycma 3.1.0

We tested our scripts on a machine with two CPUs Intel(R) Xeon(R) Silver, clocked at 2.10 GHz, and four GPUs NVidia(R) GeForce(R) RTX 2080 Ti with 12 GB each.

Warning: It is possible to utilize the software exclusively with CPU and Microsoft(R) Windows(R) as operational system. However, the performance of the software might be restricted.

Licensing

The software in this repository is licensed under the GPL-3.0. For more details on the license, please check the license file.

Software Modules

Methods

The core methods of our software are implemented in the preprocess_methods.py. The methods are divided in five classes:

  • CAE2PC: This class contains the methods for sampling 3D point clouds from computer aided design/engineering (CAD/E) models, e.g. STL meshes.

  • PC_AE: This class contains the functions to generate the parameters and corresponding Tensorflow graph for training the 3D point cloud autoencoder utilized in our research.

  • PC_VAE: This class contains the functions to generate the parameters and corresponding Tensorflow graph for training the 3D point cloud variational autoencoder utilized in our research.

  • Point2FFD: This class contains the functions to generate the parameters and corresponding Tensorflow graph for training Point2FFD, an architecture that generates polygonal meshes from input 3D point cloud representations.

  • New! Split-AE: This class contains the functions to generate the parameters and corresponding Tensorflow graph for training Split-AE, an architecture that generates content and style features of the input 3D point cloud representations for style transfer.

  • losses: Class that comprises the functions to calculate the mean-squared distance (MSD) and Kullback-Leibler divergence (KLD) on GPU for training the architectures. The Chamfer Distance (CD) is also available, however only for computation on CPU.

  • arch_training: Class with the functions to train the deep-generative models. Examples of applications are available here.

Applications

The applications of the architectures include compression of 3D point cloud data into latent representations, shape generation from latent space samples, and feature visualization. We scripted the functions for these applications in the file designapps.py in the class DesignApps. We also implemented tools for visualizing the geometric data, e.g., 3D point clouds and meshes, in the class Vis3D.

Cloning and Installation

To clone the repository through the terminal (either in Linux or GitBash), run the following command:

git clone https://github.com/HRI-EU/GDL4DesignApps.git

Otherwise, download the compressed scripts as a ZIP file from the web interface and extract the software to your directory of preference.

We recommend to install our library in a dedicated conda environment. To create and activate the environment using anaconda, utilize the following commands in a terminal window:

conda create -n designenv python=3.6 tensorflow-gpu cudatoolkit=10.1.168 matplotlib libtiff libgcc libgcc-ng cudnn cupti

conda activate designenv

After activating the environment, install the libraries that are pre-requisite for our software. To install our library (in editable mode) with pip, go to the directory where the repository was locally stored through the terminal and run the following command

pip install -e .

To test if the software was correctly installed, run the unittest.sh. The script tests if the library and modules can be imported to a Python script and tests the functionality of the main functions. If no unexpected error are reported in the terminal screen or in the file log_utest.dat, then the library was successfully installed and the functions are available for use.

How to Cite

If you think our work is interesting and utilize in your research/application, please consider citing the papers that led to the development of this library.

3D Point Cloud Autoencoder

Read our paper here.

@ARTICLE{Rios2021a,
  author={Rios, Thiago and van Stein, Bas and Bäck, Thomas and Sendhoff, Bernhard and Menzel, Stefan},
  journal={IEEE Transactions on Evolutionary Computation}, 
  title={{Multi-Task Shape Optimization Using a 3D Point Cloud Autoencoder as Unified Representation}}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TEVC.2021.3086308}}

3D Point Cloud Variational Autoencoder

Read our paper here.

@INPROCEEDINGS{Saha2020,
  author    = {Saha, Sneha and Menzel, Stefan and Minku, Leandro L. and Yao, Xin and Sendhoff, Bernhard and Wollstadt, Patricia},
  booktitle = {2020 IEEE Symposium Series on Computational Intelligence (SSCI)}, 
  title     = {{Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds}}, 
  year      = {2020},
  pages     = {1469-1477},
  doi       = {10.1109/SSCI47803.2020.9308513}
}

Feature Visualization for 3D Point Cloud Autoencoders

Read our paper here.

@INPROCEEDINGS{Rios2020a,
  author      = {Rios, Thiago and van Stein, Bas and Menzel, Stefan and Back, Thomas and Sendhoff, Bernhard and Wollstadt, Patricia},
  booktitle   = {2020 International Joint Conference on Neural Networks (IJCNN)}, 
  title       = {{Feature Visualization for 3D Point Cloud Autoencoders}}, 
  year        = {2020},
  pages       = {1-9},
  doi         = {10.1109/IJCNN48605.2020.9207326}
  }

Point2FFD

Check out out paper here

@INPROCEEDINGS{Rios2021b,
  author={Rios, Thiago and Van Stein, Bas and Bäck, Thomas and Sendhoff, Bernhard and Menzel, Stefan},
  booktitle={2021 International Conference on 3D Vision (3DV)}, 
  title={{Point2FFD: Learning Shape Representations of Simulation-Ready 3D Models for Engineering Design Optimization}}, 
  year={2021},
  volume={},
  number={},
  pages={1024-1033},
  doi={10.1109/3DV53792.2021.00110}
  }

Markdown Monster icon