This is a Pytorch implementation of the paper "GRASS: Generative Recursive Autoencoders for Shape Structures". The paper is about learning a generative model for 3D shape structures by structural encoding and decoding with Recursive Neural Networks. This code was originally written by Chenyang Zhu from Simon Fraser University and is being improved and maintained here in this repository.

Note that the current version implements only the Varational Auto-Encoder (VAE) part of the generative model. The implementation of the Generarive Adverserial Nets (GAN) part is still on-going and will be added once done. But this VAE-based model can already generate novel 3D shape structures from sampled random noises. With the GAN part, the model is expected to generate more diverse structures.

Dependancy

grass_pytorch should be run with Python 3.x. A porting to Python 2.x is provided in the folder of python2 (may not be up to date).

grass_pytorch depends on torchfold which is a pytorch tool developed by Illia Polosukhin. It is used for dynamic batching the computations in a dynamic computation graph. The computations across all nodes of all trees are batched based on their module names and dispatched to GPU for parallelization. Download and install pytorch-tools:

git clone https://github.com/nearai/pytorch-tools.git
python setup.py install

Training

python train.py

Arguments:

'--epochs' (number of epochs; default=300)
'--batch_size' (batch size; default=123 (the size of the provided training dataset is a multiple of 123))
'--show_log_every' (show training log for every X frames; default=3)
'--save_log' (save training log files)
'--save_log_every' (save training log for every X frames; default=3)
'--save_snapshot' (save snapshots of trained model)
'--save_snapshot_every' (save training log for every X frames; default=5)
'--no_plot' (don't show plots of losses)
'--no_cuda' (don't use cuda)
'--gpu' (device id of GPU to run cuda)
'--data_path' (dataset path, default='data')
'--save_path' (trained model path, default='models')

Testing

python test.py

This will sample a random noise vector of the same size as the root code. This random noise will be decoded into a tree structure of boxes and displayed using the utility functions in draw3dobb.py provided in this project.

If you use this code, please cite the following paper.

@article {li_sig17,
	title = {GRASS: Generative Recursive Autoencoders for Shape Structures},
	author = {Jun Li and Kai Xu and Siddhartha Chaudhuri and Ersin Yumer and Hao Zhang and Leonidas Guibas},
	journal = {ACM Transactions on Graphics (Proc. of SIGGRAPH 2017)},
	volume = {36},
	number = {4},
	pages = {Article No. 52},
	year = {2017}
}

This code uses the 'torchfold' in pytorch-tools developed by Illia Polosukhin.