TilinGNN: Learning to Tile with Self-Supervised Graph Neural Network (SIGGRAPH 2020)

About

The goal of our research problem is illustrated below: given a tile set (a) and a 2D region to be filled (b), we aim to produce a tiling (c) that maximally covers the interior of the given region without overlap or hole between the tile instances.

Dependencies:

This project is implemented in Python 3.7. You need to install the following packages to run our program.

• Pytorch (tested with v1.2.0): compulsory, to manipulate the tensors on GUP, and to build up the networks.
• Pytorch Geometric (tested with v1.3.2): compulsory, to build up the graph networks.
• Numpy: compulsory, to manipulate the arrays and their computations.
• Shapely (tested with v1.6.4): compulsory, for geometric computations such as collision detection.
• PyQT5: compulsory, for rendering results, and display UI interface.
• Minizinc: optional, install it only when you use IP solvers

Usage

We provide the following entry points for researchers to try our project:

• Tiling Design by UI interface: From file Tiling-GUI.py, you can use our interface to draw a tiling region and preview the tiling results interactively.
• Tiling a region of silhouette image: From file Tiling-Shape.py, you can use our pre-trained models, or IP solver, to solve a tiling problem by specifying a tiling region (from silhouette image) and a tile set.
• Training for new tile Sets: You need the following steps to train a network for a new tile set.
  1. Following the file organization of existing tile sets inside the data folder, create a new folder with new files that describe your new tile sets. After that, you need to edit the global configuration file inputs/config.py to let the system know you your new tile set.
  2. Create a superset of candidate tile placements by running file tiling/gen_complete_super_graph.py, the generated files will be stored in the folder you created in Step (1).
  3. Generate training data of random shapes by running solver/ml_solver/gen_data.py, the data will be stored in the path recorded in file inputs/config.py.
  4. Start network training by running file solver/ml_solver/network_train.py.

Note

In this program, we have a global configuration file inputs/config.py, which plays a very important role to control the behavior of the programs, such as which tile set you want to work with, the stored location of the trained networks, or how many training data you will create, etc.

Keep Improving

If you met problems or any question on this project, contact us at [haoxu@cse.cuhk.edu.hk] or [a03090@gmail.com]