shergin/malevich v1.18.0

1.18.0 (The Knife Grinder)

on GitHub

The machine-learning release. Band scales on both axes, a logarithmic
colormap, two new cells channels, and bucket-exact matrix reduction turn the
charts ML actually reads — confusion matrices, attention maps, decision
boundaries, images, loss landscapes — into grammar compositions; roc,
auc, and ewma fill out the statistical set, and a nine-chart gallery
wave plus a showcase ML corner prove the vocabulary was sufficient.
Everything is additive, and a long-standing sub-decade log-axis bug died on
the way through.

• The showcase tour grows an ML corner: a confusion matrix on band axes, a
  log-colormap attention head with a colorbar, a learned filter as rgb cells,
  1-NN decision regions with the training scatter, a momentum trajectory over
  a bucket-reduced loss landscape, seed-variance bands with EWMA smoothing,
  and a spectrogram — every panel upgrading to real pixels beside its cells
  under --features pixel, like the rest of the tour.
• The loss example's training log now credits topos by its current name —
  the library was renamed from poorgrad — and the data file moved to
  examples/data/topos_loss.csv accordingly.
• Fixed: a log axis over a range narrower than one decade no longer loses its
  data. The linear-fallback ticks of a sub-decade log range can include zero;
  the domain grew to that tick, zero has no logarithmic position, and the
  whole scale collapsed. Log domains now refuse to grow to a non-positive
  bound, and a tick without a position on its scale is dropped instead of
  drawn at a fabricated column.
• stat::ewma: debiased exponentially weighted smoothing — TensorBoard's
  scalar smoothing, early outputs unbiased instead of dragged toward zero,
  gaps passing through without disturbing the state. A scan over the ordered
  series, documented as a batch transform rather than pretending a merge law.
  Gallery gains seeds: five runs pooled into per-step quantile bands via
  the existing reducers, with the smoothed median on top.
• stat::roc and stat::auc: the classifier threshold sweep (standard step
  construction, ties grouped, one-class data returns empty rather than
  invented rates) and the trapezoid area under a polyline, gaps contributing
  no area. Batch transforms in the ecdf family — order statistics, not
  mergeable accumulators — with hand-computed fixtures. Gallery gains roc.
• Gallery: spectrogram — time × frequency power as dense Cells with a log
  frequency axis and a log colormap; the exponential chirp is a straight
  ridge. The energy is synthesized analytically: no FFT enters the crate.
• Gallery: ridgeline — distributions over training epochs as lifted KDE
  rows, painter's algorithm back to front, no camera and no new machinery:
  the TensorBoard histogram view and the honest terminal answer to a 3D
  surface.
• Gallery: calibration — a reliability diagram from stat::binned with a
  Mean reducer over 0/1 outcomes; the overconfident model's curve sags
  under the diagonal. No new API: the reducer vocabulary was sufficient.
• Gallery: landscape — a loss landscape with a momentum trajectory,
  composed entirely from existing marks (dense Cells on a log ramp, Line and
  glyph Points on top): the gradient-descent chart, no new machinery.
• Cells grids denser than the raster now reduce honestly instead of sampling:
  every screen bucket owns the cells whose centers fall inside it (adjacent
  buckets partition the centers, proven by a property test) and shows a
  reduction over all of them — Reducer::Mean by default, Cells::reduce
  to choose; Max keeps sparse spikes visible that sampling silently
  dropped. Rgb grids box-filter per channel and class grids reduce to the
  modal class with deterministic ties. 4.19 million cells reduce in ~44 ms
  on the recorded baseline (BENCHMARKS.md). Buckets owning no cell center
  keep the old center-sampling, so ordinary small grids render as before.
  Gallery gains attention-full: one million attention weights rendered
  twice, the mean pane dissolving the long-range spikes the max pane keeps.
• Cells::classes draws categorical regions: a grid of class labels colored
  through the plot's categorical Palette with a categorical legend — the
  decision-boundary chart. Labels intern in first-appearance order exactly
  like color_by; in plain output each class keeps a stable shade-ramp glyph
  and the legend swatches carry the same glyphs, so regions stay separable
  with no color at all. Gallery gains boundary, 5-NN decision regions with
  the training scatter on top.
• Cells::rgb draws a grid of direct colors — an image. Raw row-major pixel
  buffers only (decoding files stays the host's job), no colormap and no
  colorbar, honest quantization down the color ladder, and in plain output
  each pixel falls back to its luma on the shade ramp so images survive a
  pipe. With the pixel feature the grid blits at device resolution. The
  serde encoding is additive; value grids encode exactly as before. Gallery
  gains filters, an AlexNet-style Gabor bank rendered as Cells::rgb
  small multiples.
• Colormap::log() makes any ramp logarithmic: equal color steps for equal
  factors, so attention weights, gradient magnitudes, and spectral power that
  span decades stay distinguishable instead of collapsing into the low end of
  a linear ramp. Values at or below zero have no logarithmic position and
  render as gaps — the same rule log axes follow — and the colorbar places
  decade ticks logarithmically. Logarithmic and centered_at are mutually
  exclusive (validation catches the combination, including deserialized
  specs); the serde encoding stays byte-identical for existing maps. The
  gallery gains attention — token-labeled bands on both axes and a log
  MAGMA ramp.
• Scale::Bands now works on the y axis: continuous marks position y against
  band indices exactly as they do on x, and a Cells matrix maps row k onto
  band k, top-down — band 0 is the top band, so labeled matrices read in matrix
  order. Cells grids must match their band axes cell-for-band (extents do not
  apply there), Bars still require a numeric y, and Plot::y_scale no longer
  panics on Bands. Confusion matrices and attention maps are now three-line
  grammar compositions; the gallery gains confusion as the proof.