Custom activation for Saliency Maps

[trinhdhk]

I noticed that all gradient-based activation map methods implemented in MONAI implicitly assume ReLU as the target activation layer. This may very well be not the case (as in my case, it is SiLU). This leads to unreliable behaviour. Is it possible and sensible to write a custom autograd activation layer for visualising the activation map?

Thanks.

[KumoLiu]

Hi @trinhdhk, do you mean the target layer you want to generate the feature map? I think you could specify the target_layers. Here are some examples.

MONAI/monai/visualize/class_activation_maps.py

Lines 233 to 253 in 11745a6

	# densenet 2d
	from monai.networks.nets import DenseNet121
	from monai.visualize import CAM
	model_2d = DenseNet121(spatial_dims=2, in_channels=1, out_channels=3)
	cam = CAM(nn_module=model_2d, target_layers="class_layers.relu", fc_layers="class_layers.out")
	result = cam(x=torch.rand((1, 1, 48, 64)))
	# resnet 2d
	from monai.networks.nets import seresnet50
	from monai.visualize import CAM
	model_2d = seresnet50(spatial_dims=2, in_channels=3, num_classes=4)
	cam = CAM(nn_module=model_2d, target_layers="layer4", fc_layers="last_linear")
	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)

Hope it can help you, thanks!

[trinhdhk]

Thanks. No, not exactly the GradCAM. For example, here is the source code of GuidedSm

class GuidedBackpropGrad(VanillaGrad):
   

    def __call__(self, x: torch.Tensor, index: torch.Tensor | int | None = None, **kwargs) -> torch.Tensor:
        with replace_modules_temp(self.model, "relu", _GradReLU(), strict_match=False):
            return super().__call__(x, index, **kwargs)

I think the method tries to look for "relu" layers and replace it with a grad-hooked customised one. This assumingly might not work with other activation functions. Is there a way to implement a grad-hooked version of SiLU for instance?

[wyli]

thanks @trinhdhk, I'm creating a feature request from your suggestion.