I am investigating different explainability methods along with different models. I am using the library "pytorch_grad_cam" but I am still unable to know how to select the target layers so that I can get the best possible results of explainability. For example, I want to chose the best layers for DenseNet201 model. I have fine-tuned a model on my dataset, and modified the classifier layer by adding a fully connected layer. I have run an experiment using the pre-trained weights of the model and one another experiment using my trained weights. During the testing phase, my trained weights were able to score 94% while the pre-trained weights were lower than 80%, however, on the explainability, my weights look too strange while the pre-trained ones look better.
I am using
tarkget_layers = model.features[-1]
for both cases.
I also need some general guidance on how to select the best layers for explainability.