Skip to content
Docs
Inspect and Analyze
Tensor statistics

Tensor Statistics

Every view in Tensormorph that shades, colors, or summarizes a tensor - the Architecture view's magnitude-heatmap shading, the Matrix/Heatmap view, the Inspector panel's summary card - draws from the same underlying set of tile statistics, computed once and reused everywhere rather than recomputed per view.

What gets computed

For any region of a tensor (down to a single tile, up to the whole thing), Tensormorph maintains:

  • Moments - mean, variance/standard deviation, and higher moments where relevant.
  • Extremes - min, max, and absmax (the largest absolute value, which matters more than min/max alone for symmetric weight distributions).
  • Norms - L1 and squared-L2, the basis for magnitude shading and for outlier detection.
  • Special-value counts - NaN count, Inf count, and zero count (the latter doubling as the basis for sparsity measurement).
  • A log-scale histogram - the binned value distribution, which is what powers Histograms and Distributions.

Why "mergeable" matters

💡

Tile statistics are monoid-mergeable: the statistics for a coarse region can be computed directly from the statistics of its constituent finer tiles, without ever re-reading the raw values underneath them.

This is what makes level-of-detail streaming work for a trillion-parameter model: an overview of an entire layer doesn't require touching every scalar in it, only combining the already-computed statistics of its tiles. It also means a statistic is well-defined and cheap at any granularity - a single scalar, a tile, a tensor, or an entire model - using exactly the same aggregation logic throughout.

Where statistics stop being enough

Tile statistics are deliberately cheap: they're designed to be computed once during indexing and reused everywhere. Some questions need more than that - effective rank, principal components, or pairwise similarity between two tensors require actually reading more of the underlying data, not just combining precomputed summaries. Tensormorph treats those as a distinct, explicitly-requested tier of analysis rather than something that happens automatically as you navigate; see SVD, PCA and Effective Rank.

Reading statistics in the Inspector

Selecting any node - a whole tensor, a tile, a row - shows its statistics in the Inspector panel's summary card. The same numbers are queryable directly through TQL, so a question like "which layers have the largest absmax" is a query over already-computed statistics, not a scan over raw weights.

Related resources