Looking for some knowledge from a more experienced person.
I have trained two models on the “torch.hub.load” alexnet, one with pretrained=True and one with pretrained=False.
I have then applied the Grad-Cam method on the results. My interpretation is that the pretrained=True is not overfitted, it identifies the object and provide nice shapes. While the pretrained=False identifies the object, but Iooking at the gradients indicate overfitting, since it is just a splash of tiny pixels within the area of the object. No real shapes can be seen here.
Am I correct, or is there a better opinion here? (Thank you for your time.)
Pretrained = True:
Pretrained = False: