Class Activation Mapping
2 min readJan 28, 2020
Reference
Learning Deep Features for Discriminative Localization
Task
When we apply CNN to classify images, how to visualize where the CNN is paying attention to?
Method
Global Average Pooling (GAP)
GAP is proposed in Network in Network, which aims at reducing model size by replacing fully-connected layer with GAP, to compare them, let’s use the SqueezeNet 1.1 as an example,