检查Flux.jl模型中参数的总数？

Question

我知道，使用Flux.jl，我可以通过julia> Flux.params(model)来获取参数，但是输出并没有告诉我模型本身实际存在多少个总参数。是否有一个函数来检查这个，或者用一个编程的方法来计算这个？

Answer 1

正如@mcabbott在注释中提到的那样，您可以将整个模型传递给params函数，以获得每个层的总计数(sum(length, params(model)))或循环，如下所示：

julia> model = Chain(
         resnet[1:end-2],
         Dense(2048, 1000),  
         Dense(1000, 256),
         Dense(256, 2),        # we get 2048 features out, and we have 2 classes
       )
Chain(Chain(Conv((7, 7), 3=>64), MaxPool((3, 3), pad=1, stride=2), Metalhead.ResidualBlock((Conv((1, 1), 64=>64), Conv((3, 3), 64=>64), Conv((1, 1), 64=>256)), (BatchNorm(64), BatchNorm(64), BatchNorm(256)), Chain(Conv((1, 1), 64=>256), BatchNorm(256))), Metalhead.ResidualBlock((Conv((1, 1), 256=>64), Conv((3, 3), 64=>64), Conv((1, 1), 64=>256)), (BatchNorm(64), BatchNorm(64), BatchNorm(256)), identity), Metalhead.ResidualBlock((Conv((1, 1), 256=>64), Conv((3, 3), 64=>64), Conv((1, 1), 64=>256)), (BatchNorm(64), BatchNorm(64), BatchNorm(256)), identity), Metalhead.ResidualBlock((Conv((1, 1), 256=>128), Conv((3, 3), 128=>128), Conv((1, 1), 128=>512)), (BatchNorm(128), BatchNorm(128), BatchNorm(512)), Chain(Conv((1, 1), 256=>512), BatchNorm(512))), Metalhead.ResidualBlock((Conv((1, 1), 512=>128), Conv((3, 3), 128=>128), Conv((1, 1), 128=>512)), (BatchNorm(128), BatchNorm(128), BatchNorm(512)), identity), Metalhead.ResidualBlock((Conv((1, 1), 512=>128), Conv((3, 3), 128=>128), Conv((1, 1), 128=>512)), (BatchNorm(128), BatchNorm(128), BatchNorm(512)), identity), Metalhead.ResidualBlock((Conv((1, 1), 512=>128), Conv((3, 3), 128=>128), Conv((1, 1), 128=>512)), (BatchNorm(128), BatchNorm(128), BatchNorm(512)), identity), Metalhead.ResidualBlock((Conv((1, 1), 512=>256), Conv((3, 3), 256=>256), Conv((1, 1), 256=>1024)), (BatchNorm(256), BatchNorm(256), BatchNorm(1024)), Chain(Conv((1, 1), 512=>1024), BatchNorm(1024))), Metalhead.ResidualBlock((Conv((1, 1), 1024=>256), Conv((3, 3), 256=>256), Conv((1, 1), 256=>1024)), (BatchNorm(256), BatchNorm(256), BatchNorm(1024)), identity), Metalhead.ResidualBlock((Conv((1, 1), 1024=>256), Conv((3, 3), 256=>256), Conv((1, 1), 256=>1024)), (BatchNorm(256), BatchNorm(256), BatchNorm(1024)), identity), Metalhead.ResidualBlock((Conv((1, 1), 1024=>256), Conv((3, 3), 256=>256), Conv((1, 1), 256=>1024)), (BatchNorm(256), BatchNorm(256), BatchNorm(1024)), identity), Metalhead.ResidualBlock((Conv((1, 1), 1024=>256), Conv((3, 3), 256=>256), Conv((1, 1), 256=>1024)), (BatchNorm(256), BatchNorm(256), BatchNorm(1024)), identity), Metalhead.ResidualBlock((Conv((1, 1), 1024=>256), Conv((3, 3), 256=>256), Conv((1, 1), 256=>1024)), (BatchNorm(256), BatchNorm(256), BatchNorm(1024)), identity), Metalhead.ResidualBlock((Conv((1, 1), 1024=>512), Conv((3, 3), 512=>512), Conv((1, 1), 512=>2048)), (BatchNorm(512), BatchNorm(512), BatchNorm(2048)), Chain(Conv((1, 1), 1024=>2048), BatchNorm(2048))), Metalhead.ResidualBlock((Conv((1, 1), 2048=>512), Conv((3, 3), 512=>512), Conv((1, 1), 512=>2048)), (BatchNorm(512), BatchNorm(512), BatchNorm(2048)), identity), Metalhead.ResidualBlock((Conv((1, 1), 2048=>512), Conv((3, 3), 512=>512), Conv((1, 1), 512=>2048)), (BatchNorm(512), BatchNorm(512), BatchNorm(2048)), identity), MeanPool((7, 7)), #103), Dense(2048, 1000), Dense(1000, 256), Dense(256, 2))

julia> paramCount = 0
0

julia> for layer in model
           paramCount += sum(length, params(layer))
       end

julia> paramCount
25840234

在本例中，我只是在递增计数，但您可以将每个层的计数附加到数组中，例如，以单独跟踪每个层的计数。