Convolution / pooling / full connection parameter calculating
In Lenet-5 is an example:
LeNet-layer 7. 5:
. + A C1 convolution excitation
number of parameters is: (fitler + BIAS) * channel_depth
. Pooled S2 (decrease parameter, reducing over-fitting)
. A C3 vol + product excitation
core LeNet-5: use of a different number of channel convolution different feature map, in order to reduce calculation parameters (GPU performance is not high at that time)
as follows: 6. 9 + channel layer 3 layer 4 layer whole channel. 1 + connector 6 Channel
S4. pooling
not calculable parameters pooling is reduced to reduce the amount of calculation parameter
C5. convolution
extracts feature information, to do good FC fully connected
F6. fully connected
number is fully connected (input + bias) output X
output. output layer
