Skip to content

explanation_heatmaps

explanation_heatmaps

Heatmap function registry for visual explanation methods.

to_explanation_heatmaps 

to_explanation_heatmaps(channel_batch_size)

Build the registry of supported policy heatmap computations.

Parameters:

Name Type Description Default
channel_batch_size int

Number of feature channels ablated per policy forward pass for Ablation-CAM.

 required

Returns:

Type Description
dict[str, ExplanationHeatmapFunction]

Mapping from ExplanationType values to callables that compute
dict[str, ExplanationHeatmapFunction]

camera heatmaps for a policy and observation batch.
Source code in src/versatil/explainability/explanation_heatmaps.py 
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def to_explanation_heatmaps(
    channel_batch_size: int,
) -> dict[str, ExplanationHeatmapFunction]:
    """Build the registry of supported policy heatmap computations.

    Args:
        channel_batch_size: Number of feature channels ablated per policy
            forward pass for Ablation-CAM.

    Returns:
        Mapping from ``ExplanationType`` values to callables that compute
        camera heatmaps for a policy and observation batch.
    """
    return {
        ExplanationType.GRADCAM.value: partial(
            compute_gradient_maps_for_policy,
            explanation_type=ExplanationType.GRADCAM.value,
        ),
        ExplanationType.GRADCAM_PLUS_PLUS.value: partial(
            compute_gradient_maps_for_policy,
            explanation_type=ExplanationType.GRADCAM_PLUS_PLUS.value,
        ),
        ExplanationType.ABLATION_CAM.value: partial(
            compute_ablation_maps_for_policy,
            channel_batch_size=channel_batch_size,
        ),
    }