Network in Network----Global Average Pooling

1. Structure

Each feature map is a whole image for global mean pooling, and each feature map gets an output, corresponding to an output category. For example, for the CIFAR-100 classification task, the output channel of the last layer of Mlpconv can be directly set to 100, and the global average pooling is performed on each Feature Map to obtain a 100-dimensional output vector.

 

2. Function

1. GAP does not require parameters, but the fully connected layer requires a large number of parameters.

2. GAP can reduce over-fitting because there are few parameters

3. GAP summarizes spatial information, which can better adapt to spatial transformation

Guess you like

Origin http://43.154.161.224:23101/article/api/json?id=325297944&siteId=291194637