Feature map size

[xiaohythu]

Hi Nat， I found that the feature map size in your code is set to 1x1, could you tell me your reason. Natrually, feature map size should be larger than that?

[Nat-D]

Did you mean this line of code?

# x is [?, 11, 11, 32]
self.conv_feature = tf.reduce_mean(x, axis=[1,2])

This is for calculating an average activation for each feature map. High average activation would mean we detect that feature in an image regardless of its position. We use the changes in the average activation to measure whether that feature is controlled or not.