Demonstrates TNNet.WeightHistogramReport, a per-trainable-layer
weight-value histogram diagnostic.
For every layer that owns weights this helper reports min, max, mean,
population std, L2 norm, L-inf norm, sparsity (fraction of weights with
|w| < 1e-6) and an ASCII bar histogram of the weight distribution
across Bins bins over [-MaxAbs, +MaxAbs] (per-layer scaling so each
layer fills its own range). Biases are excluded; the closing line
reports total trainable weight count and trainable-layer count.
The example builds a 2-hidden-layer ReLU MLP and trains it briefly on a
synthetic y = ||x|| regression task. It prints the histogram twice
(before and after training) and a one-line max-abs summary, so the shift
in distribution shape and range is visible.
cd examples/WeightHistogramReport
lazbuild WeightHistogramReport.lpi
../../bin/x86_64-linux/bin/WeightHistogramReport
Runs well under 30 seconds, pure CPU, no datasets required.