What does the global pool do?

Question

I recently discovered the global_pooling flag in the Pooling layer in caffe, but I could not find it in the documentation here ( Layer Catalog) nor here ( Doxygen union doc ).

Is there a simple explanation for this simple explanation compared to the usual behavior of a pool-layer?

Answer 1

With a global pool, reduces dimensionality from 3D to 1D. Therefore, the global pooling of results 1 is responsible for each function map. This can be the maximum or average, or any other join operation that you use.

It is often used at the end of the backend of the convolutional neural network to get a shape that works with dense layers. Therefore, it is not necessary to use a sealant.

Answer 2

Convoys can work with any size of image input (which is large enough). However, if you have a fully connected layer at the end, this layer requires a fixed input size. Therefore, for a complete network, a fixed input image size is required.

However, you can remove the fully connected layer and just work with convolutional layers. You can make a convolution layer at the end that has the same number of filters as the classes. But you need one value for each class that indicates the probability of this class. Therefore, you apply a merge filter across the remaining functions map. This association, therefore, is “global” because it is always as large as necessary. Conversely, regular merge layers have a fixed size (for example, 2x2 or 3x3).

This is a general concept. You can also find the global pool in other libraries, for example. Lasagne . If you need good literature, I recommend reading Network In Network .

Answer 3

If you are looking for information on flags / coffee options, it is better to look at them in the comments '$CAFFE_ROOT/src/caffe/proto/caffe.proto' .
For the 'global_pooling' parameter, the comment says :

 // If global_pooling then it will pool over the size of the bottom by doing // kernel_h = bottom->height and kernel_w = bottom->width 

For more information on caffe layers, see this help page .