Squeeze-and-Excitation Networks, a new architecture building block proposed by WMW team at the ILSVRC2017 challenges(http://image-net.org/challenges/LSVRC/2017/results#team)
Figure 1: Inception and resnet with SE module.
For more information, please refer to this article(http://www.haik8.com/p/jxz4p.html)
You can add followed module to any net structure.
...
layer {
...
top: "xxx"
...
}
layer {
name: "se2_1/pool"
type: "Pooling"
bottom: "xxx"
top: "se2_1/pool"
pooling_param {
pool: AVE
global_pooling: true
}
}
layer {
name: "se2_1/ip1"
type: "InnerProduct"
bottom: "se2_1/pool"
top: "se2_1/ip1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 2
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu_se2_1/ip1"
type: "ReLU"
bottom: "se2_1/ip1"
top: "se2_1/ip1"
}
layer {
name: "se2_1/ip2"
type: "InnerProduct"
bottom: "se2_1/ip1"
top: "se2_1/ip2"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 32
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "sigmoid_se2_1/ip2"
type: "Sigmoid"
bottom: "se2_1/ip2"
top: "se2_1/ip2"
}
layer {
name: "se2_1/scale"
type: "Scale"
bottom: "xxx"
bottom: "se2_1/ip2"
top: "se2_1/scale"
scale_param {
axis: 0
}
}
layer {
...
bottom: "se2_1/scale"
...
}
...