Figures¶

torch_world.plots draws the figures used in the notebooks, in one visual system: a fixed categorical order, one sequential hue for time, hairline axes and 2px lines. It needs matplotlib:

pip install "torch-world[examples]"

Every function returns a matplotlib Figure, which notebooks display (at retina resolution) and scripts can save with fig.savefig(...).

function	draws
`images(x, titles, ncols)`	a grid of `(N, 3, H, W)` images
`masks(image, grid, layers)`	an image beside one panel per mask, tinting the kept patches
`curves(series, smooth)`	training curves: raw values under a moving average, labeled at the end
`trajectories(paths, goals)`	small multiples of planar paths toward goals, with a success radius
`distances(runs, radius)`	distance to goal per step: median over the interquartile band, per method
`bars(groups, baseline)`	horizontal grouped bars with value labels
`answers(x, questions, predictions, truths)`	question-answer cards marked right or wrong
`probes(true, pred, names)`	linear-probe scatter per target, with R²

from torch_world import plots

log = {"prediction": [], "SIGReg": []}
...  # training
plots.curves(log, smooth=100)

For custom figures in the same style, use the context manager and the palette:

with plots.style() as plt:
    fig, ax = plt.subplots()
    ax.plot(values, color=plots.SERIES[0])