3dgan-chainer

Chainer implementation of 3D Generative Adversarial Network.

Result

Some good samples generated chairs. (50epoch)

python generate_samples.py result/trained_models/Generator_50epoch.npz <save direcotry> <num to be generated>

Requirements

• chainer(2.0.1)
• scipy
• scikit-image
• h5py

pip install scipy scikit-image h5py

Optional

• If you want to plot voxel, latest matplotlib is needed.

  The 3D voxel / volumetric plot function is included only in unreleased version of matplotlib as of Oct 19, 2017

  • matplotlib 2.1.0+323.ge6448bafc

  pip install git+https://github.com/matplotlib/matplotlib
  

Dataset

I used ShapeNet-v2 dataset. Training script support .binbox or .h5 extension.

Describe your dataset path to DATASET_PATH in train.py.

.binvox

Just use .binvox files in ShapeNet-v2.

.h5

Assuming that .h5 has { 'data': <np.array, shape (64, 64, 64)> }. If you want to convert .binvox into .h5, use binvox_to_h5.py script.

Usage

Training

python train.py

Generation

python generate_samples.py <model_file> <save_dir> <num samples>

Visualization

If you have .binvox file, using simple-voxel-viewer is easy way.

Or visualize with matplotlib,

python visualize.py <binvox file or directory>