GradCAM

Grad-CAM class-discriminative localisation

This example demonstrates TNNet.GradCAMReport, an implementation of Grad-CAM (Gradient-weighted Class Activation Mapping, Selvaraju et al., 2017). Grad-CAM produces a coarse, class-discriminative heatmap by weighting a convolution layer's feature maps with the gradient of the predicted logit and keeping only the positive evidence:

alpha_k = mean_xy( d y_c / d A^k_xy )          (global-average-pooled gradient)
L_xy    = ReLU( sum_k alpha_k * A^k_xy )        (weighted, rectified feature sum)

It complements the fine, input-pixel attribution of SaliencyReport: saliency tells you which pixels the network reacted to, Grad-CAM tells you which region of the conv feature map drove the class decision. This is the resolution/locality trade-off: Grad-CAM is coarse (it lives at a conv layer's feature-map grid, here 8x8) but class-discriminative; saliency is fine (full input resolution) but noisier and not inherently class-localised.

The task

A tiny CNN (forked from the SaliencyReport demo) learns to separate two classes of 8x8x2 images:

• class 0: a bright 3x3 blob in the top-left of channel 0
• class 1: a bright 3x3 blob in the bottom-right of channel 1

After a short training loop (well under a minute on CPU, no memory blow-up) the program explains one class-0 prediction two ways on the same sample:

1. GradCAMReport — coarse Grad-CAM localisation at the deepest spatial conv layer
2. SaliencyReport — fine pixel-space saliency / SmoothGrad / Integrated Gradients

Both are printed as ASCII heatmaps so the example is fully self-contained (no image files).

Built-in self-check

The example parses the coarse Grad-CAM peak cell out of the report, and asserts it lands inside the class-0 (top-left) region of the input. It prints PASS / FAIL and Halt(1)s on failure, so it doubles as a regression gate. GradCAMReport is forward-only: it runs Compute + a one-hot Backpropagate to read activations and gradients but never calls UpdateWeights, so the trained weights are left untouched.

Running

lazbuild examples/GradCAM/GradCAM.lpi
./bin/<arch>/bin/GradCAM

Expected tail of the output:

Self-check: Grad-CAM coarse peak cell = (0,1)
PASS: Grad-CAM peak falls inside the class-0 (top-left) region.

API

class function TNNet.GradCAMReport(
  NN: TNNet;                  // trained classifier
  Probe: TNNetVolume;         // input sample (already shaped for the net)
  ConvLayerIdx: integer = -1; // target conv layer (-1 = deepest spatial conv)
  ForcedClass: integer = -1   // class to attribute (-1 = predicted argmax)
): string;                    // ASCII report (coarse map + nearest-upsampled overlay)