I know that I should apply Dropout on training and testing, get several results for prediction and so on. But, considering a pretrained model, how should I add Dropout layers?
- on the fully-connected layers;
- after every convolutional layer.
I’ve read some papers and implementations where one applies Dropout at fully-connected layers only, using a pretrained model; however, when using a custom model, usually one adds Dropout after every conv layer.
Using Dropout on the fully-connected layers only can be considered an actual uncertainty estimation of the model?