Signal and Distribution Editor
The Signal and Distribution view is Tensormorph's 1D view family - for reading a value as a curve or a distribution along a single axis, rather than as a spatial grid.
The 1D sub-modes
| Sub-mode | Shows |
|---|---|
| Raw Value Plot | A single row, column, or channel's values plotted in order - the most direct 1D read of a tensor slice |
| Distribution | The value distribution as a log-scale histogram - see Histograms and Distributions |
| Layer Profile | A chosen statistic (mean, norm, sparsity) plotted across every layer of a model, for spotting which layer is the outlier |
| Expert Profile | The same idea as Layer Profile, scoped to the experts within one MoE layer - see Mixture-of-Experts Models |
| Spectrum | The singular-value spectrum of a selected matrix - see SVD, PCA and Effective Rank |
| Delta | The 1D counterpart to a weight diff: a row/column/channel's values in one revision plotted against the same slice in another |
Why a dedicated 1D family, not just a "zoomed-in" 2D view
A row or channel plotted as a curve reveals shape-of-distribution and outlier-position information that's genuinely harder to read off a 2D heatmap - a heatmap encodes value as color, which is good for spatial pattern but comparatively weak for reading exact relative magnitude across many points at once. Layer Profile and Expert Profile in particular have no natural 2D or 3D analogue: they're inherently "one number per layer/expert, plotted in order," which is a 1D question from the start.
Playback and scrubbing
Where a Layer Profile or Expert Profile is plotted across an ordered axis (layer index, training checkpoint, morph-interpolation step), the view supports scrubbing through that axis with a playhead - the same interaction as scrubbing through runtime trace captures, applied to a static ordered axis instead of a time axis.
Related resources
- Matrix and Heatmap Editor - the 2D counterpart for a full tile rather than a single row or column.
- Histograms and Distributions - the Distribution sub-mode in full detail.
- Tensor Statistics - the underlying statistics every sub-mode above plots.