Implementation of a 3D shape generative model based on deep convolutional generative adversarial nets (DCGAN) with techniques of improved-gan.

Experimental results on ShapeNetCore dataset are shown below. For training the networks, I used all 3D models in ShapeNetCore.

This is an application for visualizing linear interpolation and saving generated data as binvox. You can run this application with the following command:

$ python application.py

I strongly recommend running the app on GPU because it is very slow on CPU.

To train the networks, you need to install three python packages.

• binvox-rw-py
• numpy
• TensorFlow

The following python packages are required for running the application. If you are using anaconda, you can easily install VTK5 and PyQt4 (or they may already be installed). I show installation commands with conda for VTK5 and PyQt4.

• VTK5

$ conda install -c anaconda vtk=5.10.1

• PyQt4

$ conda install -c anaconda pyqt=4.11.4

• QDarkStyleSheet

$ pip install qdarkstyle

1. Install the python packages above.
2. Download the code from GitHub:

$ git clone https://github.com/maxorange/voxel-dcgan.git
$ cd voxel-dcgan

3. Specify dataset path and model path in config.py:

...
dataset_path = "path/to/dataset/*.binvox"
params_path = "path/to/model"
...

4. Train the networks:

$ python train.py

5. Generate data:

$ python visualize.py
or
$ python application.py

More details are here.