Moving shapes dataset

A simple 2d temporal data generator written in PyTorch. Generates frames with a configurable amount of shapes of different shapes and velocities.

Example usage:

import torchvision

if torch.cuda.is_available():
    device = 'cuda'
else:
    device = 'cpu'

shapes = MovingShapes(
    box_size = 16,
    surface_size = 32,
    object_size = 3,
    batch_size = 3,
    num_shapes = 3,
    speed_int = (1,3),
    device=device)

for i in range(20):
    surface = shapes.step()
    torchvision.transforms.ToPILImage()(surface[0].detach().cpu().view(32,32)).show()

Outputs:

Parameters

• surface_size is the size of the surface the shapes are drawn on.
• box_size determines the boundaries in which the random vertex positions of the shapes are chosen.
• object_size is the amount of lines, or rather vertices, each shape has.
• batch_size is the amount of examples are calculated at once. For efficient processing on GPU's, use a high batch size and divide it up in smaller batches after.
• num_objects determines the amount of shapes drawn on the surface in each example.
• speed_int is the interval from which uniformly the velocity for each shape is chosen. A speed of 1 means it moves 1 pixel per frame.