Introduction

This is an implementation of a simple 2d physics simulator usable from the pytorch framework.

The main function is

torch.ByteTensor flatland.generate_sequence(bool pulling,
                                            int nb_sequences,
                                            int nb_images,
                                            int image_height, int image_width,
                                            int nb_shapes,
                                            bool random_shape_size, bool random_colors)

where

pulling indicates if one of the shape should be pulled upward,
nb_sequences is the number of sequences to generate,
nb_images is the number of frames per sequence,
image_height and image_width is the individual image size,
nb_shapes is the number of rectangular shapes in the simulation,
random_shape_size states if the shapes should be of different sizes,
random_colors states if they should be of different colors.

The returned ByteTensor has five dimensions:

Sequence index
Image index
Channel (3 since it is RGB)
Pixel row
Pixel col

Warning: For improper parameters values (e.g. too many shapes, too small images), the simulation may be stuck.

Installation and test

make -j -k
./test-flatland.py