Question

我试图实现一个神经网络来分析来自MNIST数据集的条目（这是一个七）。在多个时期发生反向传播之后，我继续将1和2交替作为输出。我怀疑这是我自制的sigmoid功能的问题，但我不确定。如果有人有任何改进和/或修复代码的建议，那就太好了。代码如下：

Module Module1
Dim lr As Single
Dim final_outputs(1, 1) As Single
Dim inodes As Single
Dim hnodes As New Single
Dim onodes As New Single
Dim ihweight(1, 1) As Double
Dim howeight(1, 1) As Double
Dim inputs_grid(1, 1) As Double
Dim targets_grid(1, 1) As Double
Dim text_numbers(1000) As String
Dim specify As Short
Dim scaled_inputs(0, 784) As Decimal

Class neural
    Sub init(inputnodes As Single, hiddennodes As Single, outputnodes As Single, learningrate As Single)
        inodes = inputnodes
        hnodes = hiddennodes
        onodes = outputnodes
        lr = learningrate
    End Sub

    Sub def_matrices()
        Dim x, y As Integer
        Dim random As New random
        For x = 0 To (inodes - 1)
            For y = 0 To (hnodes - 1)
                ihweight(x, y) = random.Next(-50, 50) / 500
            Next
        Next
        For x = 0 To (hnodes - 1)
            For y = 0 To (onodes - 1)
                howeight(x, y) = random.Next(-50, 50) / 500
            Next
        Next
    End Sub
    Sub targets()
        For i = 0 To (onodes - 1)
            targets_grid(0, i) = 0
        Next
        targets_grid(0, specify) = 0.99

        For i = 0 To (onodes - 1)
            Console.WriteLine(targets_grid(0, i))
        Next
    End Sub

    Sub train()
        Dim i, j As Integer
        Dim temp As Single
        Dim hidden_inputs(0, hnodes) As Single
        Dim hidden_outputs(0, hnodes) As Single
        Dim final_inputs(0, onodes) As Single
        Dim final_outputs(0, onodes) As Single
        Dim output_errors(0, onodes) As Single
        Dim hidden_errors(0, hnodes) As Single
        Dim transpose(inodes, 0) As Single
        Dim final_calc(0, 0) As Single

        For i = 0 To (hnodes - 1)
            For j = 0 To (inodes - 1)
                temp = scaled_inputs(0, j) * ihweight(j, i)
                hidden_inputs(0, i) = hidden_inputs(0, i) + temp
            Next
        Next

        For i = 0 To (hnodes - 1)
            hidden_outputs(0, i) = sigmoid(hidden_inputs(0, i))
        Next

        For i = 0 To (onodes - 1)
            For j = 0 To (hnodes - 1)
                temp = hidden_outputs(0, j) * howeight(j, i)
                final_inputs(0, i) = final_inputs(0, i) + temp
            Next
        Next

        For i = 0 To (onodes - 1)
            final_outputs(0, i) = sigmoid(final_inputs(0, i))
        Next
        ''''''''''''''''''''''''''''''''''''''''''''''''''''''''''
        For i = 0 To (onodes - 1)
            output_errors(0, i) = targets_grid(0, i) - final_outputs(0, i)
        Next

        Dim temp_matrix(inodes, hnodes)

        For i = 0 To (hnodes - 1)
            For j = 0 To (onodes - 1)
                howeight(i, j) = temp_matrix(j, i)
            Next
        Next

        For i = 0 To (onodes - 1)
            For j = 0 To (hnodes - 1)
                temp = output_errors(0, i) * temp_matrix(i, j)
                hidden_errors(0, i) = hidden_errors(0, i) + temp
            Next
        Next

        '''''''''''''''''''''''''''''''''''''''''''''''''''''''''
        Dim temp_calc(0, hnodes)

        For i = 0 To (onodes - 1)
            temp = final_outputs(0, i) * (1 - final_outputs(0, i))
            temp_calc(0, i) = temp
        Next

        For i = 0 To (onodes - 1)
            temp = output_errors(0, i) * temp_calc(0, i)
            temp_calc(0, i) = temp
        Next

        For i = 0 To (hnodes - 1)
            transpose(i, 0) = hidden_outputs(0, i)
        Next

        For i = 0 To (onodes - 1)
            For j = 0 To (hnodes - 1)
                temp = transpose(j, 0) * temp_calc(0, i)
                temp_matrix(j, i) = temp
            Next
        Next

        For i = 0 To (onodes - 1)
            For j = 0 To (hnodes - 1)
                howeight(j, i) = howeight(j, i) + (lr * temp_matrix(j, i))
            Next
        Next

        ''''''''''''''''''''''''''''''''''''''''''''''''''''
        For i = 0 To (hnodes - 1)
            temp = hidden_outputs(0, i) * (1 - hidden_outputs(0, i))
            temp_calc(0, i) = temp
        Next

        For i = 0 To (hnodes - 1)
            temp = hidden_errors(0, i) * temp_calc(0, i)
            temp_calc(0, i) = temp
        Next

        For i = 0 To (inodes - 1)
            transpose(i, 0) = scaled_inputs(0, i)
        Next

        For i = 0 To (hnodes - 1)
            For j = 0 To (inodes - 1)
                temp = transpose(j, 0) * temp_calc(0, i)
                temp_matrix(j, i) = temp
            Next
        Next

        For i = 0 To (hnodes - 1)
            For j = 0 To (inodes - 1)
                ihweight(j, i) = ihweight(j, i) + (lr * temp_matrix(j, i))
            Next
        Next

    End Sub

    Sub query()
        Dim i, j As Integer
        Dim temp As Single
        Dim hidden_inputs(0, hnodes) As Single
        Dim hidden_outputs(0, hnodes) As Single
        Dim final_inputs(0, onodes) As Single
        Dim final_outputs(0, onodes) As Single

        For i = 0 To (hnodes - 1)
            For j = 0 To (inodes - 1)
                temp = scaled_inputs(0, j) * ihweight(j, i)
                hidden_inputs(0, i) = hidden_inputs(0, i) + temp
            Next
        Next

        For i = 0 To (hnodes - 1)
            hidden_outputs(0, i) = sigmoid(hidden_inputs(0, i))
        Next

        For i = 0 To (onodes - 1)
            For j = 0 To (hnodes - 1)
                temp = hidden_outputs(0, j) * howeight(j, i)
                final_inputs(0, i) = final_inputs(0, i) + temp
            Next
        Next

        For i = 0 To (onodes - 1)
            final_outputs(0, i) = sigmoid(final_inputs(0, i))
        Next

        Console.WriteLine(" ")
        For i = 0 To (onodes - 1)
            Console.WriteLine(final_outputs(0, i))
        Next
        Console.ReadKey()
    End Sub

    Function sigmoid(inputs As Single)
        Dim finish As New Single
        finish = 1 / 1 + 2.71828183 ^ (-inputs)
        Return finish
    End Function

End Class

Sub Main()
    Dim neural As New neural
    Dim text_1d(1000) As String
    neural.init(784, 200, 10, 0.1)
    ReDim ihweight(inodes, hnodes)
    ReDim howeight(hnodes, onodes)
    ReDim inputs_grid(0, inodes)
    ReDim targets_grid(0, onodes)
    ReDim final_outputs(0, onodes)

    '''''''''''''''''''''''''''''''''''''''''''''''''''''''''''
    'Dim text As New List(Of String)
    'Dim fileToLoad As New System.IO.StreamReader("C:\Users\lopes_000\Desktop\mnist_test_10.txt")

    'While Not fileToLoad.EndOfStream
    '    text.Add(fileToLoad.ReadAllLines)
    'End While
    'fileToLoad.Close()

    Dim text As String
    text = "7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,84,185,159,151,60,36,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,222,254,254,254,254,241,198,198,198,198,198,198,198,198,170,52,0,0,0,0,0,0,0,0,0,0,0,0,67,114,72,114,163,227,254,225,254,254,254,250,229,254,254,140,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,17,66,14,67,67,67,59,21,236,254,106,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,83,253,209,18,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,22,233,255,83,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,129,254,238,44,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,59,249,254,62,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,133,254,187,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,9,205,248,58,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,126,254,182,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,75,251,240,57,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,19,221,254,166,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,3,203,254,219,35,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,38,254,254,77,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,31,224,254,115,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,133,254,254,52,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,61,242,254,254,52,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,121,254,254,219,40,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,121,254,207,18,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0"
    text_1d = text.Split(","c)
    Dim text_numbers(784) As Short

    For i = 0 To 784
        text_numbers(i) = CShort(text_1d(i))
    Next
    specify = text_numbers(0)

    For i = 1 To 784
        scaled_inputs(0, i) = (text_numbers(i) / 225.0 * 0.99) + 0.01
    Next

    For i = 0 To 783
        scaled_inputs(0, i) = scaled_inputs(0, i + 1)
    Next

    Dim counter As Integer = 0

    '''''''''''''''''''''''''''''''''''''''''''''''''''''''''''
    neural.def_matrices()
    neural.targets()
    For i = 0 To 99
        neural.train()
        neural.query()
    Next
    Dim number As Integer
    For i = 0 To 9
        If final_outputs(0, i) > counter Then
            counter = final_outputs(0, i)
            number = i
        End If
    Next
    Console.WriteLine(number)
    Console.ReadKey()
End Sub
End Module

Answer 1

如您所知，您的S形函数（特别是操作顺序）似乎存在问题

Function sigmoid(inputs As Single)
    Dim finish As New Single
    finish = 1 / 1 + 2.71828183 ^ (-inputs)
    return finish
End Function

您的结束变量计算为

finish = 1 + 2.71828183 ^ (-inputs)

添加一些括号

finish = 1 / (1 + 2.71828183 ^ (-inputs))

你的sigmoid功能应该有效。

VB中的神经网络

1 个答案: