plot_anchors

plot_anchors(anchors, boxes=None, log=False, *, model=None, split='train')

Fitted anchors against the box shapes they were fitted to.

Two ways in, the same two R has, where plot_anchors() takes a fit. Give a fitted DetectionEngine and it asks the engine for both clouds in one call, colouring the boxes by class; or give the anchors and the shapes yourself.

Handing it the engine is the safer of the two and not merely the shorter: widths from one place and anchors from another is how a figure comes to show boxes in different places from where the anchors were fitted to them, which looks like a bad fit and is a bug. box_shapes returns both as fractions of the letterboxed input, which is the convention the anchors were fitted in - taking the source image instead is what had a coverage of 0.67 quoted for two years against a true 0.65.

Parameters

Name	Type	Description	Default
anchors	Any	a fitted DetectionEngine, or one group of (width, height) pairs per grid, as fractions of the input.	required
boxes	np.ndarray | None	the shapes themselves, n x 2 of widths and heights, drawn underneath. Refused together with an engine, which already has them.	None
log	bool	both axes on a log scale, off by default as in R. Detection datasets span a wide range of sizes - a BCCD platelet is about two fifths the side of a red cell - and a log scale spreads the classes a linear one crowds into the corner.	False
model	str | None	which detector, when an engine is given. The default is the first.	None
split	str	which split’s boxes, when an engine is given. Worth drawing for a split the anchors were not fitted on: it is the only way to see a class the anchors have nothing near.	'train'

Returns

Name	Type	Description
	Figure	A matplotlib Figure.

Raises

Name	Type	Description
	PlatypusError	if both an engine and boxes are given.

figure = plot_anchors([[(0.1, 0.1), (0.2, 0.3)], [(0.4, 0.5)]]) len(figure.axes[0].collections), figure.axes[0].get_yscale() (1, ‘linear’)