Class GradCAM

monai/visualize/class_activation_maps.py:319–384

Computes Gradient-weighted Class Activation Mapping (Grad-CAM). This implementation is based on: Selvaraju et al., Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization, https://arxiv.org/abs/1610.02391 Examples .. code-block:: python

class GradCAM(CAMBase):
    """
    Computes Gradient-weighted Class Activation Mapping (Grad-CAM).
    This implementation is based on:

        Selvaraju et al., Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization,
        https://arxiv.org/abs/1610.02391

    Examples

    .. code-block:: python

        import torch

        # densenet 2d
        from monai.networks.nets import DenseNet121
        from monai.visualize import GradCAM

        model_2d = DenseNet121(spatial_dims=2, in_channels=1, out_channels=3)
        cam = GradCAM(nn_module=model_2d, target_layers="class_layers.relu")
        result = cam(x=torch.rand((1, 1, 48, 64)))

        # resnet 2d
        from monai.networks.nets import seresnet50
        from monai.visualize import GradCAM

        model_2d = seresnet50(spatial_dims=2, in_channels=3, num_classes=4)
        cam = GradCAM(nn_module=model_2d, target_layers="layer4")
        result = cam(x=torch.rand((2, 3, 48, 64)))

    N.B.: To help select the target layer, it may be useful to list all layers:

    .. code-block:: python

        for name, _ in model.named_modules(): print(name)

    See Also:

        - :py:class:`monai.visualize.class_activation_maps.CAM`

    """

    def compute_map(self, x, class_idx=None, retain_graph=False, layer_idx=-1, **kwargs):  # type: ignore[override]
        _, acti, grad = self.nn_module(x, class_idx=class_idx, retain_graph=retain_graph, **kwargs)
        acti, grad = acti[layer_idx], grad[layer_idx]
        b, c, *spatial = grad.shape
        weights = grad.view(b, c, -1).mean(2).view(b, c, *[1] * len(spatial))
        acti_map = (weights * acti).sum(1, keepdim=True)
        return F.relu(acti_map)

    def __call__(self, x, class_idx=None, layer_idx=-1, retain_graph=False, **kwargs):  # type: ignore[override]
        """
        Compute the activation map with upsampling and postprocessing.

        Args:
            x: input tensor, shape must be compatible with `nn_module`.
            class_idx: index of the class to be visualized. Default to argmax(logits)
            layer_idx: index of the target layer if there are multiple target layers. Defaults to -1.