请问如何将每一层神经网络和第一层神经网络的输出相加呢，如图

学科相关 机器学习
1

P}OC5`IEG3

S%~R3KDFDO_CHS0)BA_{99
S%~R3KDFDO_CHS0)BA_{99827×455 40.1 KB

第一张图是我想构建的网络结构，第二张图是我写的程序，但是这个函数是无法调用出参数的

# Struct to define model
function m(x)
    pqsize = 72
    vsize = 36
    layer0 = Dense(pqsize, 52, relu)
    layer1 = Dense(52, 32, relu)
    z1 = 
    passthrough1 = Dense(pqsize, 32, relu)
    layer2 = Dense(32, 21, relu)
    passthrough2 = Dense(pqsize, 21, relu)
    layer3 = Dense(21, 30, relu)
    passthrough3 = Dense(pqsize, 30, relu)
    layer4 = Dense(30, vsize, relu)
    return layer4(passthrough3(x) + layer3(passthrough2(x) +
                         layer2(passthrough1(x) + layer1(layer0(x)))))
end

这是我写的程序，希望好心的小可爱能给我指个方向呜呜呜

2

P}OC5`IEG3

3

这个是我想实现的~

4

简单来说你这里基本结构是对的，但是参数没写对

struct PassThroughBlock
    forward
    passthrough
end

# 告诉Flux这是一个Flux兼容的网络层，这个也许可以不写，但不太确定
Flux.@functor PassThroughBlock

# 先定义一个方便的构造函数
function PassThroughBlock(Ns::Tuple; activation = relu)
    Ls = [Dense(Ns[1], Ns[2], activation)]
    Ps = []

    for (n_in, n_out) in zip(Ns[2:end-1], Ns[3:end])
        push!(Ls, Dense(n_in, n_out, activation))
        push!(Ps, Dense(Ns[1], n_out, activation))
    end

    return PassThroughBlock(Chain(Ls...),  Chain(Ps...))
end

# 接下来定义怎么样进行前向传播
function (block::PassThroughBlock)(x)
    Ls = block.forward
    Ps = block.passthrough

    z = Ls[1](x)
    for (l, p) in zip(Ls[2:end], Ps)
        z = l(z) + p(x)
    end
    return z
end

# 测试
x = ones(72, 10)
Ns = (72, 52, 32, 21, 30, 36)
block = PassThroughBlock(Ns)
block(x) # array of size (36, 10)
2 个赞
5

呜呜呜呜，谢谢大神，Julia社区真好，我会好好努力的呜呜呜

1 个赞