Skip to content
Docs
Editors
Matrix and Heatmap Editor

Matrix and Heatmap Editor

The Matrix and Heatmap view is Tensormorph's 2D view family - the tool for reading a specific 2D slice of a tensor precisely, at the cost of the whole-tensor spatial overview the Tensor Volume view gives you instead.

Shading modes

A single 2D grid, with the shading mode determining what each cell's color means:

ModeShows
SignedValue with sign - positive/negative distinguished by color, not just magnitude
MagnitudeAbsolute value only
SparsityThe sparsity mask - see Sparsity
QuantizationQuantization granularity groups - see Dtype and Quantization
Diff|A − B| between two aligned tensors - see Weight Diff
CorrelationPairwise correlation, e.g. between experts - see Expert Similarity
SVDRank-truncated reconstruction or singular-value contribution - see SVD, PCA and Effective Rank
OutlierDeviation from a local statistical baseline - see Outliers and Anomalies

The diverging color convention

💡

Wherever a shading mode has a natural positive/negative or above/below-baseline meaning (Signed, Diff), Tensormorph uses one consistent diverging palette: cyan for positive, red for negative, dark for unchanged/zero. This convention holds across every view that shows signed data, not just this one.

Consistency here matters specifically because you regularly compare a Matrix/Heatmap panel against the Architecture view's magnitude shading or the Distribution view's mean marker in the same session - a color meaning something different in each would defeat the point of shading anything at all.

Cell-level precision

Unlike the Tensor Volume view, hovering a cell here shows its exact value immediately (not just an approximate color read), and clicking opens it in the Scalar Inspector for the full metadata. Zooming further subdivides the grid rather than blending cells together, so precision doesn't degrade as you zoom in - only the visible extent shrinks.

Related resources