L^c = ReLU(sum_k alpha^c_k * A^k), where alpha^c_k = mean_ij(d(y^c)/d(A^k_ij)) and A is the last OpType::Conv node's activation. More...

#include <grad_cam.hpp>

Public Member Functions
Attribution	explain (ExplainerContext &ctx, const Tensor &input, int64_t target_index, DeviceBackend *backend) const
	Computes the Grad-CAM map for one target class index.

Detailed Description

L^c = ReLU(sum_k alpha^c_k * A^k), where alpha^c_k = mean_ij(d(y^c)/d(A^k_ij)) and A is the last OpType::Conv node's activation.

Note: Finds the target layer via ComputationGraph::nodes_by_op_type(OpType::Conv)'s last entry – the theory file's stated default ("the last convolutional layer before the classifier head"), found by op-type tag per charter Part 2 SS3, never by layer name.; Reads the target node's own activation/gradient as ComputationGraph tagged it – in this codebase Conv and its following ReLU are separate graph nodes, so "the last conv layer" here means the pre-ReLU conv output specifically. This is a legitimate placement (Grad-CAM's formula doesn't require a post-ReLU activation), just worth noting for a reader used to "conv block" framing elsewhere.; A pure graph walker built entirely against ExplainerContext's public interface – no core (Tensor/ComputationGraph/Autograd/Module) changes needed.; Device-generic: the CAM weighting runs on host copies of the target node's activation/gradient; the seed and the CAM are uploaded beside the network's tensors.

Member Function Documentation

Attribution pulsatrix::GradCAM::explain	(	ExplainerContext &	ctx,
		const Tensor &	input,
		int64_t	target_index,
		DeviceBackend *	backend
	)		const

inline

Computes the Grad-CAM map for one target class index.

Parameters

ctx	Context to run the forward/backward pass through. Its graph must contain at least one OpType::Conv node.
input	Input to explain.
target_index	Which output element (class score) to attribute (0-based, flat index into the network's final output).
backend	Backend to allocate the one-hot seed and result tensors through.

Returns: An Attribution with method "grad_cam", values = the N x H x W CAM batch (same spatial size as the target conv layer's feature maps, one map per batch example), and metadata recording the target index and which node was used.

Note: Assumes a rank-2 (N, num_classes) network output and a rank-4 (N, channels, H, W) target conv layer activation – migrated by campaign_exai_dl_library_batch_dimension_support from the original rank-1/ rank-3 assumptions. Same single-shared-target_index scope boundary as Saliency's identical migration.

The documentation for this class was generated from the following file: