Python 在其他编程语言上使用Matlab中导出的神经网络权重

我在使用python复制matlab中创建的神经网络时遇到问题。它是一个

{9,8,4}

网络。下面分别是matlab和python的原始输出

0.00187283763854096 0.00280257304094145 0.00709416898379967 0.00474275971385824 0.000545071722266366
0.0520122170317888  0.0402746073491970  0.0179208146529717  0.0245726107168336  0.230693355244371
0.430695009441386   0.434492291029203   0.410151021812136   0.416871471927059   0.469873849186641
0.562954025662924   0.539410486293765   0.666336481449288   0.637779009735872   0.284564488176231

[1.0, -1.0, -0.6875955603907775, -0.9999999426232321]
[1.0, -1.0, 0.5569364789701737, -0.9994593106654553]
[1.0, -1.0, 0.5022468075847347, -0.999780120038859]
[1.0, -1.0, 0.4924691499951816, -0.9997110849203137]
[1.0, -1.0, 0.5945295051094253, -0.9991584098381949]

我使用

net2.IW{1}

，

net2.LW{2}

获得输入和分层权重。得出的偏差如下：

net2.b{1}

和

net2.b{2}

没有使用偏见，我得到了一些看起来很接近的东西：

[-0.6296705512038354, 0.9890465283687858, 0.1368924025968622, 0.5426776395855755]
[-0.05171165478856995, 0.2973298654798701, 0.02897695903082293, 0.0499820714219222]
[-0.10046933055782481, 0.40531232885083035, 0.033067381241777244, 0.06585830703439044]
[0.03167268710874907, 0.5485036035542894, 0.10579223668518502, 0.015475934153332364]
[0.006502829360007152, 0.22928662468119648, 0.03788967208701787, 0.012868192806301859]

因此，我认为问题可能在于偏见；不过我不太确定

从Matlab获取权重的Python实现

def sigmoid(x):
    return math.tanh(x)

def NN(inputs, bias1, bias2):
    wsum=[sum([(x*y) for x,y in zip(inputs[0],row)])for row in inputweights]
    wsbias=[(x + y) for x,y in zip(wsum,bias1)]
    inputactivation=[sigmoid(k) for k in wsbias]
    wsoutput=[sum([(x*y) for x,y in zip(inputactivtion,row)])for row in     hiddenweights]
    wsbias2=[(x + y) for x,y in zip(wsoutput,bias2)]
    outputactivation=[sigmoid(k) for k in wsbias2] 
    return 'output' outputactivation

我非常感谢任何有效的解决方案。 下面是获得的输入和分层权重以及输入和分层偏差

IW=[[-9.1964,   -2.3015,    0.2493,    3.3648,   -2.6015,   -0.0795,  -11.2356,    4.6861,-0.8360],
    [6.0201,   -1.8708,    2.7844,    0.2419,   -1.1808,   -8.6800,    5.8519,   -5.2958,    5.3233],
    [0.8597,    0.8644,   -0.6913,   -0.0397,    0.0619,    0.4506,    1.0687,    0.4090,   -0.2874],
    [2.9459,    3.2596,    2.2859,    1.1933,    2.9675,   -9.6017,    3.5893,    1.4808,   -7.5311],
   [-0.1533,   -1.4806,   -2.3748,    0.8059,   -0.5502,   -1.0447,   -0.5920,   -1.1667,   -1.1447],
    [4.7185,   -9.2097,    1.1001,   -0.0173,    1.4929,    0.3884,    3.7674,    6.3459,   -4.2845],
  [-16.4031,    8.1351,    2.0689,    2.1267,    6.2093,   -8.3875,  -15.8493,   -0.6096,    2.9214],
    [1.7329,    0.1797,   0.1500,    9.1616,   -1.7226,    0.9479,    3.2542,  -24.4003,   -4.2790]]

LW=[[-18.5985,   12.2366,  -0.8833,   -1.6382,    4.6281,    8.1221,  -23.7587,   -0.8589],
   [12.0462,  -11.5464,    6.9612,  -10.8562,   -7.0647,    5.6653,   16.2527,   -7.6119],
   [12.4176,    0.9808,    0.7650,  -2.9434,   -0.2765,   -3.0689,   -3.1528,   3.0389],
    [5.7570,   7.7584,  -6.9550,   -2.3679,   -1.4884,  -11.0668,    2.6764,   26.5427]]

bias1=[-1.7639, -1.2599, -0.7560, 0.2520,-0.2520,0.7560, -1.2599, -1.7639]
bias2= [0.2129,-8.1812, 0.0202,4.4512]

我的投入

[[0.0, 0.0, 0.0414444125526, 0.0, 0.0, 0.00670501464516, 0.0, 0.0, 0.0313140652051], [0.0, 0.0, 0.0, 1.0]]
[[0.0, 0.0, 0.00398243636152, 0.0, 0.0, 0.000863557858377, 0.0, 0.0, 0.00356406423776], [0.0, 0.0, 0.0, 1.0]]
[[0.0, 0.0, 0.00440892765754, 0.0, 0.0, 0.000725737283104, 0.0, 0.0, 0.00543503005753], [0.0, 0.0, 0.0, 1.0]]
[[0.0, 0.0, 0.00565322288091, 0.0, 0.0, 0.00236630383341, 0.0, 0.0, 0.00642911490856], [0.0, 0.0, 0.0, 1.0]]
[[0.0, 0.0, 0.00250332223564, 0.0, 0.0, 0.000926998841251, 0.0, 0.0, 0.00241792804103], [0.0, 0.0, 0.0, 1.0]]

---

感谢您的建议。

偏差有时作为权重矩阵的第一列或最后一列添加，有时作为自己的向量添加。但这并不依赖于语言，而是依赖于库。你能说出图书馆的名字吗？还有一些关于如何在Python中创建网络/权重的代码（显示您在何处使用了先前提取的值。另外，您确定使用的是相同的传递/激活函数吗？@Neil Slater，谢谢您的回答。我使用了nprtool和tansig transfor函数。@NeilSlater，我使用了从matlab导出的权重和偏差在python中实现。您的偏差是否存在于输入中？