Hey SY and YW, to use the pretrained GoogLeNet, I built the model with GoogleNet by 4e layer, followed by one conv layer and 3 dropout fully connected layer.

To run the code

python googlenet_drop.py

In googlenet_drop.py file,

• to freeze the model, you can set the argument freeze to True
• there are two images that are saved in img directory. The first one is 4e layer (to see that the front layers are freezed), and the second one is the feature map by the output of the last conv layer.

As I've mentioned before, the last feature map is used for generating CAM. However, after training the model, the last feature map seems to have sparse map, and this leads to a wrong CAM.