nngraph的问题解决

在这篇
Torch7入门续集补充— nngraph包的使用 中已经初步了解nngraph， 但是在实际应用中，可能会出现一些问题。这里只知道原因，解决方法比较ugly。

node declared on [[c]]:-1_ does not connect to gmodule output

在很多情况下，我们可能只想根据特定条件，截取预定义好的部分网络。比如：

function defineG_unet_swap(input_nc,output_nc,ngf, swapLayer64, bottlneck)

  local net_G=nil


    local e1 = - nn.SpatialConvolution(input_nc, ngf, 4, 4, 2, 2, 1, 1)
    -- input is (ngf) x 128 x 128
    local e2 = e1 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf, ngf * 2, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 2)
    -- input is (ngf * 2) x 64 x 64
    local e3 = e2 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 4) 
    -- input is (ngf * 4) x 32 x 32
    local e4 = e3 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 4, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8)
    -- input is (ngf * 8) x 16 x 16
    local e5 = e4 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8)
    -- input is (ngf * 8) x 8 x 8
    local e6 = e5 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8)
    -- input is (ngf * 8) x 4 x 4
    local e7 = e6 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8)
    -- input is (ngf * 8) x 2 x 2
    local e8 = e7 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, bottlneck, 4, 4, 2, 2, 1, 1) -- nn.SpatialBatchNormalization(ngf * 8)
    -- input is (ngf * 8) x 1 x 1

    local d1_ = e8 - nn.ReLU(true) - nn.SpatialFullConvolution(bottlneck, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8) -- nn.Dropout(0.5)
    -- input is (ngf * 8) x 2 x 2
    local d1 = {d1_,e7} - nn.JoinTable(2)
    local d2_ = d1 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8) -- nn.Dropout(0.5)
    -- input is (ngf * 8) x 4 x 4
    local d2 = {d2_,e6} - nn.JoinTable(2)
    local d3_ = d2 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8) -- nn.Dropout(0.5)
    -- input is (ngf * 8) x 8 x 8
    local d3 = {d3_,e5} - nn.JoinTable(2)
    local d4_ = d3 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 8)
    -- input is (ngf * 8) x 16 x 16
    local d4 = {d4_,e4} - nn.JoinTable(2)
    local d5_ = d4 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 4)
    -- input is (ngf * 4) x 32 x 32
    local d5 = {d5_,e3} - nn.JoinTable(2)
    local d6_ = d5 - nn.ReLU(true)  - nn.SpatialFullConvolution(ngf * 4 * 2, ngf * 2, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 2)
    -- input is (ngf * 2) x 64 x 64
    local d6 = {d6_,e2} - nn.JoinTable(2)
    local d7_ = d6 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2 * 2, ngf, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf)
    -- input is (ngf) x128 x 128
    local d7 = {d7_,e1} - nn.JoinTable(2)
    local d8 = d7 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2, output_nc, 4, 4, 2, 2, 1, 1)
    -- input is (nc) x 256 x 256
    local o1 = d8 - nn.Tanh()     

    -- 这里是条件，比如
    if end_layer == 4 then
        netG = nn.gModule({e1},{e2 - nn.LeakyReLU(0.2, true)})
    elseif  ...
         ...
    end

    return netG
end

根据传入的end_layer来确定不同的截取的网络。然而这样写就会出现上面的问题。
查看出错地方：

         local errStr =
            "node declared on %s does not connect to gmodule output"
         assert(successorIsInGraph,
                string.format(errStr, successor.data.annotations._debugLabel))

可以看到这 successorIsInGraph….这就尴尬了，这说明e2的有些后继结点不在定义的图内。可以看到e3你只定义了，但是不用，也会出错。所以你只能每个结点的定义都放在条件语句中了。类似：

    local e1 = - nn.SpatialConvolution(input_nc, ngf, 4, 4, 2, 2, 1, 1) -- 1

    if end_layer == 1 then
        local e=e1 - nn.LeakyReLU(0.2, true)
        return nn.gModule({e1},{e})
    end

    -- input is (ngf) x 128 x 128
    local e2 = e1 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf, ngf * 2, 4, 4, 2, 2, 1, 1) - nn.SpatialBatchNormalization(ngf * 2) --2 3 4

    if end_layer == 4 then
        return nn.gModule({e1},{e2})
    end
    ...