Where should I set the target_layer when applying Gradcam?

[jaehyeok-jeon]

Hello, I am trying to apply gradcam to the RT-DETR_pytorch model.

Existing models such as resnet attach a hook to the act layer of layer_4(the last layer feature extracotor), the feature extractor, and calculate grad_cam through the corresponding feature and gradient values.

In transformer models such as rt-detr, I am curious as to whether I should hook the backbone/res_layers/3/blocks/2/act layer, which is the last layer of the backbone, or whether I should hook the three input_projs of the decoder.

If you need to hook a different layer than what I thought, I would appreciate it if you could tell me where that layer is.