I(x, y, no of channels) is the image, and Fi(x, y, no of filters ) is the feature map at some layer 'i'. Given the architecture of a Convolutional Neural Network like VGGNet and a feature map after a certain layer Fi, is there an efficient way to find which pixels of the input image I, that contribute to a location in the feature map? I will want to implement this in python.
Asked
Active
Viewed 152 times
4
-
I believe what you are looking for is a way to visualize the features which have been learned by an individual convolutional filter. If so you might want to checkout this post https://towardsdatascience.com/how-to-visualize-convolutional-features-in-40-lines-of-code-70b7d87b0030 For more try searching for `convolutional feature visualization` – Meto May 10 '20 at 11:27
-
try grad-cam https://stackoverflow.com/questions/61336309/how-to-implement-cam-without-visualize-cam-in-this-code/61336761#61336761 – Zabir Al Nazi May 10 '20 at 11:27
-
I have been maintaining kind of a list of tools used for interpretability and diagnostics, which I try to keep up to date; have a look, I guess you'll find several options: https://stackoverflow.com/questions/52391871/predictive-analytics-why-factor-model-interpretability/52392344#52392344 – desertnaut May 10 '20 at 16:07