search

facebookresearch / ToMe

A method to increase the speed and lower the memory footprint of existing vision transformers.
Other
931 stars 67 forks source link

Is there any way to generate heatmap with ToMe #14

Open Nightwatch-Fox11 opened 1 year ago

Nightwatch-Fox11 commented 1 year ago

Hello, Thanks for this amazing work!

I'm wondering if you've tried generating heatmap of the attention weight before, like Grad-CAM. For example, after I get the attention weight of each attention block and the corresponding sourcematrix, is there any way to generate heatmap with ToMelike the following image: image

Looking forward to your reply :)

Best