This is a modified version of minGPT for point cloud processing. Please look at the original source for more detailed information about minGPT. The Jupyter Notebook play_PC.ipynb is used to train minGPT model with point clouds. All details are in the notebook with comments and markdowns.
The "chair" class from ShapeNet Part dataset is used after randomly downsampling 3746 models to 1024 points. Any point cloud dataset in (Models x Points x 3) format can be used.
The code is tested with PyTorch v1.5.0. Training takes around 3 hours on a single RTX 2070.
| Dataset |
|---|
 |
| After discretization |
 |
| Generated samples |
 |
 |