Skip to content

score_cam

muppet.explainers.score_cam

Score-CAM (Score-weighted Class Activation Mapping) explainer for CNNs.

This module implements Score-CAM, a gradient-free method for generating class activation maps that overcomes limitations of gradient-based CAM approaches. Instead of using gradients to weight feature maps, Score-CAM evaluates the contribution of each activation map by using it as a mask and measuring the resulting change in the model's confidence.

MUPPET Component Integration
  • Explorer: CAMExplorer - extracts feature maps from the model's last convolutional layer
  • Perturbator: SetToZeroPerturbator - applies feature map-based masking to input
  • Attributor: ClassScoreAttributor - evaluates model confidence on masked inputs
  • Aggregator: WeightedSumAggregator - combines feature maps weighted by their scores with ReLU post-processing

Classes:

References

Wang, Haofan, et al. "Score-CAM: Score-weighted visual explanations for convolutional neural networks." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2020. https://arxiv.org/pdf/1910.01279.pdf

Classes

ScoreCAMExplainer 
ScoreCAMExplainer(model, convention='constructive')

Bases: MuppetExplainer

ScoreCAM (Score-weighted Class Activation Mapping) explainer implementation.

Implements ScoreCAM method that generates visual explanations by weighting feature maps based on their contribution scores to the final prediction, providing parameter-free class activation mappings.

Score-CAM addresses several issues with traditional gradient-based methods: - Eliminates gradient saturation problems - Provides more reliable importance scores - Removes dependence on gradient computations - Offers better visual explanation quality

The method works by: 1. Extracting feature maps from the last convolutional layer 2. Using each feature map as a mask to perturb the input image 3. Evaluating the masked input to get confidence scores 4. Computing a weighted combination of feature maps using these scores

This approach provides a more direct measurement of each feature map's contribution to the final prediction, leading to more accurate and interpretable visualizations.

Initialize the Score-CAM explainer for CNN explanation.

Parameters:

  • model (Module) –

    The convolutional model to explain its predictions.

  • convention (Union[AttributionConvention, str], default: 'constructive' ) –

    choose if the explainer finds important features by identifying features that destroy (destructive) efficiently the model's prediction from the input, or by identifying features that build (constructive) efficiently the model's response from a completly perturbed input

Source code in muppet/explainers/score_cam.py 
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def __init__(
    self,
    model: torch.nn.Module,
    convention: Union[AttributionConvention, str] = "constructive",
) -> None:
    """Initialize the Score-CAM explainer for CNN explanation.

    Args:
        model (torch.nn.Module): The convolutional model to explain its predictions.
        convention: choose if the explainer finds important features by identifying
            features that destroy (destructive) efficiently the model's prediction
            from the input, or by identifying features that build (constructive)
            efficiently the model's response from a completly perturbed input
    """
    # Parameters

    explorer = CAMExplorer(model=model)
    perturbator = SetToZeroPerturbator()
    attributor = ClassScoreAttributor(convention=convention)
    aggregator = WeightedSumAggregator(
        convention=convention,
        post_proc=lambda x: torch.nn.functional.relu(x),
    )

    # Initialize the explainer with these modules
    super().__init__(
        model=model,
        explorer=explorer,
        perturbator=perturbator,
        attributor=attributor,
        aggregator=aggregator,
    )