使用从Neuronet（R）获得的权重手动计算输出

Question

我想手动计算神经网络模型的输出（预测），其中权重是使用R中的神经网络函数获得的。

我有以下代码，

library(neuralnet)
library(datasets.load)

nn<- neuralnet(formula = dist ~ mag , data =  attenu, hidden = c(1,2),act.fct = "logistic" ,stepmax=1e6)

我得到以下结果：

> nn$weights
[[1]]
[[1]][[1]]
          [,1]
[1,] -9.219761
[2,]  1.347328

[[1]][[2]]
          [,1]      [,2]
[1,] -396.1329 -330.0593
[2,]  982.5444  537.1123

[[1]][[3]]
         [,1]
[1,] 23.00682
[2,] 57.26966
[3,] 78.06727

相关模型输出如下：

预测输出如下：

> nn$net.result
[[1]]
            [,1]
  [1,]  80.27648
  [2,] 158.34375
  [3,] 158.34375
  [4,] 158.34375
  [5,] 158.34375
  [6,] 158.34375
  [7,] 158.34375
  [8,] 158.34375
  [9,] 158.34375
 [10,] 158.34375
 [11,] 158.34375
 [12,]  23.00682
 [13,]  23.00682
 [14,]  23.00682
 [15,]  23.00682
 [16,]  23.00682
 [17,]  23.00682
 [18,]  23.00682
 [19,]  23.00682
 [20,]  23.00682
 [21,]  23.00682
 [22,]  80.27647
 [23,]  80.27647
 [24,]  80.27647
 [25,]  80.27647
 [26,]  80.27647
 [27,]  80.27647
 [28,]  80.27647
 [29,]  80.27647
 [30,]  80.27647
 [31,]  80.27647
 [32,]  80.27647
 [33,]  23.00682
 [34,]  23.00682
 [35,]  23.00682
 [36,]  23.00682
 [37,]  23.00682
 [38,]  23.00682
 [39,]  23.00682
 [40,]  80.27647
 [41,]  80.27647
 [42,]  80.27647
 [43,]  80.27647
 [44,]  80.27647
 [45,]  80.27647
 [46,]  80.27647
 [47,]  80.27647
 [48,]  80.27647
 [49,]  80.27647
 [50,]  80.27647
 [51,]  80.27647
 [52,]  80.27647
 [53,]  80.27647
 [54,]  80.27647
 [55,]  80.27647
 [56,]  80.27647
 [57,]  80.27647
 [58,]  80.27647
 [59,]  80.27647
 [60,]  80.27647
 [61,]  80.27647
 [62,]  23.00682
 [63,] 158.34375
 [64,] 158.34375
 [65,] 158.34375
 [66,]  23.00682
 [67,]  23.00682
 [68,]  23.00682
 [69,]  23.00682
 [70,]  23.00682
 [71,]  23.00682
 [72,]  23.00682
 [73,]  23.00682
 [74,]  23.00682
 [75,]  23.00682
 [76,]  23.00682
 [77,]  23.00682
 [78,]  23.00682
 [79,]  23.00682
 [80,] 158.34375
 [81,] 158.34375
 [82,] 158.34375
 [83,]  23.00682
 [84,]  23.00682
 [85,]  23.00682
 [86,]  23.00682
 [87,]  23.00682
 [88,]  23.00682
 [89,]  23.00682
 [90,]  23.00682
 [91,]  23.00682
 [92,]  23.00682
 [93,]  23.00682
 [94,]  23.00682
 [95,]  23.00682
 [96,]  23.00682
 [97,]  23.00682
 [98,]  23.00682
 [99,]  23.00682
[100,]  23.00682
[101,]  23.00682
[102,]  23.00682
[103,]  23.00682
[104,]  23.00682
[105,]  23.00682
[106,]  23.00682
[107,]  23.00682
[108,]  23.00682
[109,]  23.00682
[110,]  23.00682
[111,]  23.00682
[112,]  23.00682
[113,]  23.00682
[114,]  23.00682
[115,]  23.00682
[116,]  23.00682
[117,]  23.00682
[118,]  23.00682
[119,]  23.00682
[120,]  23.00682
[121,]  23.00682
[122,]  23.00682
[123,]  23.00682
[124,]  23.00682
[125,]  23.00682
[126,]  23.00682
[127,]  23.00682
[128,]  23.00682
[129,]  23.00682
[130,]  23.00682
[131,]  23.00682
[132,]  23.00682
[133,]  23.00682
[134,]  23.00682
[135,]  23.00682
[136,]  23.00682
[137,]  23.00682
[138,]  23.00682
[139,]  23.00682
[140,]  23.00682
[141,]  23.00682
[142,]  23.00682
[143,]  23.00682
[144,]  23.00682
[145,]  23.00682
[146,]  23.00682
[147,]  23.00682
[148,]  23.00682
[149,]  23.00682
[150,]  23.00682
[151,]  23.00682
[152,]  23.00682
[153,]  23.00682
[154,]  23.00682
[155,]  23.00682
[156,]  23.00682
[157,]  23.00682
[158,]  23.00682
[159,]  23.00682
[160,]  23.00682
[161,]  23.00682
[162,]  23.00682
[163,]  23.00682
[164,]  23.00682
[165,]  23.00682
[166,]  23.00682
[167,]  23.00682
[168,]  23.00682
[169,]  23.00682
[170,]  23.00682
[171,]  23.00682
[172,]  23.00682
[173,]  23.00682
[174,]  23.00682
[175,]  23.00682
[176,]  23.00682
[177,]  23.00682
[178,]  23.00682
[179,]  23.00682
[180,]  23.00682
[181,]  23.00682
[182,]  23.00682

但是，使用获得的权重并进行计算，我无法设法获得Neuronet（来自R）获得的输出。

根据结果，对于第一个元素（mag = 7），我将输出计算为：

x=7
u=-9.21976+1.34733*x
 =-9.21976+1.34733*7
 =0.21

a1=1/(1+EXP(-(396.1329+982.5444*u)))=0
a2=1/(1+EXP(-(330.059+537.112*u)))=0

output=23.0068+57.2697*a1+78.0673*a2
      =23.0068

但是，神经网络函数的结果为80.276。

我对结果感到困惑。我无法获得相同的结果。

我将很高兴获得任何帮助。非常感谢。

Answer 1

激活功能适用于每个单元，包括偏置。在您的示例中，输出80.27648的计算公式为：

act <- function(x) 1/(1 + exp(-x))
u <-  act(-9.21976 + 1.34733*x)
a1 <- act(-396.1329 + 982.5444*u)
a2 <- act(-330.059 + 537.112*u)
output <- 23.0068 + 57.2697*a1 + 78.0673*a2