How can we retrieve attention mask from the deep learning model?

Question

Hi I want to visualize what my model has learned and want to get a attention mask from my model. May I ask how can I do this?

Below is my desire output.

From paper [New perspectives on plant disease characterization based on deep learning]

Answer 1

The method that they apply is cited in the paper:

6.1.1. Qualitative results [...] we used one of the visualization methods presented in (Yosinski et al., 2015) [...].

From a quick glimpse, it looks like Yosinski et. al. visualize the magnitude of pixels of a particular feature map. The authors of your referenced paper provide more detail how exactly they produced the results:

More precisely, we started by plotting the position of the neuron with the highest activation for all the features maps extracted from the last convolutional layer, and from there we accumulated the first 30 dominant activations and matched them to the original image. [Lee et. al.]