search

kailaix / ADCME.jl

Automatic Differentiation Library for Computational and Mathematical Engineering
https://kailaix.github.io/ADCME.jl/latest/
MIT License
286 stars 57 forks source link

you can't label a fc network with biases #68

Open yue-SJTU opened 3 years ago

yue-SJTU commented 3 years ago

maybe I'm wrong,but I found you may not mark the network with biases,such as fc(x, config, θ,"label"),which would arise error

kailaix commented 3 years ago

When you provide θ, you do need "label" because the deep neural network is uniquely identified by θ.

"label" is useful when you build a neural network by specifying weights and biases implicitly. For example

y1 = fc(x, config) 
y2 = fc(x, config)

y1 and y2 are outputs of the same DNN.

y1 = fc(x, config,"label1") 
y2 = fc(x, config,"label2")

y1 and y2 are outputs of two different DNNs.