SaliencyReport

SaliencyReport

Tiny example for TNNet.SaliencyReport, the forward+backward input-attribution (saliency) diagnostic.

Given a trained classifier and a single probe sample (a TNNetVolume input), the report picks the predicted class c = argmax(f(x)) and computes three flavours of per-input-element attribution, printing each as a compact ASCII heatmap over the input plane (one block per channel, one cell per pixel, ~10 intensity buckets " .:-=+*#%@", brightest = highest |attr|):

• (a) vanilla input-gradient saliency |d logit_c / d x| — one forward pass, then one backward pass with the final-layer error set to the one-hot e_c, so the gradient that reaches the input layer is exactly d logit_c / d x;
• (b) SmoothGrad — (a) averaged over N noisy copies x + eta, eta ~ N(0, sigma^2), sigma = SmoothNoiseFrac * (max(x) - min(x)) (denoises the raw gradient);
• (c) Integrated Gradients — the straight-line path integral from a zero baseline to x in K steps: IG_i ~= (x_i - 0) * (1/K) * sum_{k=1..K} d logit_c / d x_i at (k/K)*x.

Per channel it also prints the total attribution mass for each variant and the top-K most-attributing pixel coordinates of the vanilla map.

Built-in correctness check: the IG completeness gap

Integrated Gradients satisfies the completeness axiom sum_i IG_i ~= logit_c(x) - logit_c(0). The report prints the completeness gap |sum(IG) - (logit_c(x) - logit_c(0))| (absolute and relative). A small relative gap means the path integration is faithful — this is the one-number sanity/regression check for the report.

Pure CPU; reuses the existing forward/backward path. No optimiser step is applied, so the inspected network's weights are left untouched.

What this demo shows

On a synthetic 8x8x2 two-class image task (class 0 = bright 3x3 blob in the top-left of channel 0; class 1 = bright blob in the bottom-right of channel 1, plus background noise) it trains a small conv classifier (Conv -> MaxPool -> FC -> SoftMax), then runs the report on:

1. a clean, correctly-classified class-0 probe — the three heatmaps should concentrate around the discriminative top-left/channel-0 blob, and the IG completeness gap is small;
2. a noisier (harder) class-1 probe — attribution spreads out more.

The raw input is printed as its own ASCII heatmap above each report so the attribution maps can be eyeballed against the actual input content.

Build & run

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

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