EquivarianceReport

EquivarianceReport

Tiny example for TNNet.EquivarianceReport, the forward-only input-symmetry (invariance / equivariance) diagnostic.

Given a network and a probe batch of inputs, the report measures how the forward output reacts to a fixed menu of input-side symmetry transforms and prints, per transform:

• the invariance error = mean over the probe batch of ||f(T(x)) - f(x)||_2 / ||f(x)||_2 (0 means the model ignores the transform, i.e. is invariant to it);
• the top-1 agreement rate mean(argmax(f(T(x))) == argmax(f(x))) (meaningful for classifier-shaped outputs);
• a 10-bin ASCII histogram of the per-sample invariance error so outliers are visible;
• a one-line verdict: invariant (err < InvariantTol, default 1e-3), approximately invariant (err < ApproxTol, default 1e-1) or sensitive.

The default transform menu for image-shaped inputs is TNNetFlipX (horizontal mirror), TNNetFlipY (vertical mirror), TNNetReverseChannels (channel reversal) and a 1-channel TNNetRoll (depth roll). Each T(x) is produced by a tiny Input -> Transform forward-only wrapper net; no backward pass is run and the inspected network's weights are never touched.

What this demo shows

On a tiny synthetic 8x8x3 3-class image task it builds and trains two classifiers, then prints the report for each:

1. NET A — a plain conv classifier (Conv -> MaxPool -> FC -> SoftMax). It has no built-in spatial symmetry, so it is flip-sensitive: the FlipX / FlipY rows report a large invariance error (verdict sensitive).
2. NET B — Input -> TNNetAvgChannel -> FC -> SoftMax. A global per-channel spatial average is unchanged by any spatial permutation, so this net is FlipX- and FlipY-invariant by construction: the FlipX / FlipY rows report ~0 invariance error (verdict invariant). It is still sensitive to the channel permutations (ReverseChannels / Roll), which is visible in the contrast.

This is the built-in correctness check: a net that is invariant to a transform by construction reads ~0 invariance error on that transform's row.

Build & run

cd examples/EquivarianceReport
lazbuild EquivarianceReport.lpi
../../bin/x86_64-linux/bin/EquivarianceReport

Pure CPU, no dataset download. Total runtime is well under a minute.