SVGSampler

This is a small project to quickly create 2D multiclass-datasets for machine learning applications.

You can install it using pip:

pip install svg_sampler

Then you can use the function sample_from_svg to sample datapoints from the filled in paths/objects of the svg.

Examples:

from svg_sampler import sample_from_svg
from matplotlib import pyplot as plt

X, y = sample_from_svg(path_to_file, 5000, normalize=True,
                       sample_setting="based_on_area", overlap_mode="upper_only")
plt.scatter(X[:, 0], X[:, 1], c=y)
plt.show()

Documentation

sample_from_svg

sample_from_svg(path, total_samples, sample_setting="equal_over_classes", overlap_mode="all", normalize=False, *, seed)

Parameters

• path (str): Path to the SVG file.
• total_samples (int): Total number of points to sample.
• sample_setting (str): Sampling mode. Options:
  • "equal_over_classes" – Union shapes by fill and sample equally.
  • "equal_over_shapes" – Sample equally from each shape.
  • "based_on_area" – Allocate samples proportional to the shape's area.
• overlap_mode (str): Overlap handling mode. Options:
  • "all" – Sample from all overlapping shapes.
  • "upper_only" – Only sample from the top (upper) shape in overlapping regions.
• normalize (bool): If True, normalize the sampled coordinates per axis to [0, 1].
• seed (int): Seed for random number generation.

Returns

• X (ndarray): Array of shape (N, 2) with (x, y) coordinates.
• y (ndarray): Numeric class labels for each sample point.

Dependencies

numpy
shapely
svgpathtools
matplotlib