You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Global Average and Max Pooling are used as an alternative to FC layers.
they are very handy for vizualization. what quiver can do is show the convolution layers output that are input to GAP but GAP calculates average activation for a single channel of previous convolution.
it would be good to show this number alongside the visualized output of convolution layer if the next layer is GAP or GMP.
The text was updated successfully, but these errors were encountered:
And furthermore there is yet another architecture where GAP is followed by a dense layer before given to softmax layer: http://cnnlocalization.csail.mit.edu/
supporting this as well would be really cool.
Actually those are the types of architectures that allow a built-in localization of classified item
Global Average and Max Pooling are used as an alternative to FC layers.
they are very handy for vizualization. what quiver can do is show the convolution layers output that are input to GAP but GAP calculates average activation for a single channel of previous convolution.
it would be good to show this number alongside the visualized output of convolution layer if the next layer is GAP or GMP.
The text was updated successfully, but these errors were encountered: