How is Grad-CAM picture drawed?

[beierzhu]

How is Grad-CAM picture drawed?

[machengcheng2016]

Understood. Much appreciate!

[BeierZhu]

Dear Chengcheng Ma:

Thank you for your interest in ProGrad.

For Figure 2, we use the last convolutional layer of CLIP-ResNet50 (clip_model.visual.layer4[2]) to draw the Grad-CAM.
And adapt the code from this link Grad-CAM

Best

[machengcheng2016]

Thanks for your reply.
I think the clip_model.visual.layer4[2] always remain fixed, because both CoOp and ProGrad only train the prompt about the text model and never change the parameters of visual model. So why are the GradCAM pictures different?

[BeierZhu]

Because the classification head generated by prompt changes

[BeierZhu]

CAM activation depends on the derivative of the logits which is a function of classification head