Created by: guoshengCS

How to do global_pooling which can pool over the whole feature map and return a scalar value for each feature map without indicating pool_size and stride. With this maybe the configurations of inception-resnet-v2 and inception-v4 can support different input sizes.