Has anyone tried using EigenGradCAM for Class Activation Mapping of YOLOv5?

[BaTteREg]

During the process of adding back - propagation gradients, I found that the output of the model in the BaseCAM function, self.outputs, has a size of (25200, 85). When continuing with the gradient calculation, how should I select the output tensors to calculate the loss and complete the back - propagation?

       if self.uses_gradients:
            self.model.zero_grad()
            # for param in self.model.parameters(): # 设置requires_grad=True
            #     if not param.requires_grad:
            #         param.requires_grad = True
            for name, param in self.model.named_parameters():
                print(f"Parameter name: {name}, requires_grad: {param.requires_grad}")
            loss = sum([target(output) for target, output in zip(targets, outputs)])
            loss.backward(retain_graph=True)
            if 'hpu' in str(self.device):
                self.__htcore.mark_step()

outputs, has a size of (25200, 85).

[BaTteREg]

May I ask whether the output of the YOLO model needs to go through Non-Maximum Suppression (NMS) processing here?