在训练具有三重态损失的暹罗网络时，批处理规范层权重/参数使NaN变为

Question

参考，我注意到“ NaN”仅存在于“ batch_norm”层中。为什么会发生这种情况？

培训是否存在一些问题-例如“消失梯度”或“梯度爆炸”？还是NaN仅是偏差的占位符？

（只是一个前提，我正在尝试训练一个具有三重损失的暹罗网络，如此处Keras Model Weights for some layers become all "NaNs"（架构和损失）所述）

此处列出的样品重量-

Yes batch_normalization_11                                                  
[array([-0.08248743, -0.19357444,  0.06189843, -0.03510274,  0.00484655,
   -0.00720728,  0.04206916, -0.08482166,  0.04828389, -0.07205907, 
   -0.06575502, -0.08964766, -0.05986991, -0.03338304, -0.11991593, 
    0.02784944,  0.0775785 , -0.15297122,  0.07871342,  0.12960596, 
    0.0105051 ,  0.00689312, -0.03168776, -0.01650888, -0.01664159, 
   -0.0392631 ,  0.01485326,  0.01542591,  0.04976993,  0.03047125, 
   -0.06873457, -0.00606309, -0.01566394, -0.00729234,  0.02118564, 
   -0.02180263,  0.04011146, -0.0738162 , -0.06278081,  0.00753566, 
   -0.06096083, -0.04877877, -0.01127802,  0.04446675, -0.02236561, 
    0.01035857, -0.02206145, -0.03292053, -0.01618384, -0.03100098, 
    0.0676478 , -0.10707849, -0.01150786,  0.04351562, -0.01791094, 
   -0.00431769, -0.0049645 ,  0.04900745,  0.03982552, -0.03878992, 
    0.00659816, -0.02593656, -0.02570266, -0.12747312,  0.0494753 , 
   -0.01329002, -0.00264208,  0.03973205, -0.100038  , -0.01836757, 
    0.0236165 ,  0.02873622,  0.02456198, -0.02947933, -0.03124348, 
   -0.0517611 , -0.03143985, -0.04440191, -0.00843053, -0.00296288, 
    0.01514868,  0.01892053,  0.01062289,  0.00799673,  0.17570479, 
   -0.03454979,  0.04009602,  0.0713095 , -0.12413061,  0.00034242, 
   -0.06435894,  0.0845928 ,  0.03802553, -0.1751569 ,  0.01174091, 
   -0.08430453], dtype=float32), array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, 
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,  
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,  
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,  
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,  
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,  
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, 
   nan, nan, nan, nan, nan], dtype=float32), array([138.71233 , 143.4597  ,  52.366314, 239.61154 , 105.63001 ,
   169.46217 , 168.73875 , 190.50938 , 188.52414 ,  75.47664 ,           
   165.98553 , 169.36244 , 152.27237 ,  66.28865 ,  71.25762 ,           
   194.51697 , 100.27181 , 120.084465, 252.51112 , 108.817085,           
   127.83531 , 184.69302 ,  52.401028, 172.47392 , 178.41951 ,           
   158.10571 , 107.729454, 239.42154 , 153.06093 , 210.01903 ,           
   171.95633 , 172.42186 ,  94.53224 , 171.9558  ,  98.768486,           
   107.27287 , 167.34947 , 121.79589 , 139.91867 , 185.08607 ,           
   113.509705, 156.48535 , 173.54575 ,  74.16201 , 168.05954 ,           
   159.63252 ,  96.90468 , 175.31023 , 111.47005 , 169.68703 ,           
   146.08734 ,  76.32107 , 150.7184  , 197.71414 , 111.61184 ,           
   159.9483  , 153.95746 , 201.43568 , 141.50627 , 141.01636 ,           
   126.97432 ,  42.3665  , 165.11559 , 165.68098 , 160.95749 ,           
   176.41151 , 147.87408 , 119.34812 ,  51.771393, 105.82823 ,           
   182.31734 , 219.16847 , 136.30508 , 225.81941 , 158.1635  ,           
   193.53983 , 140.47253 , 142.7074  , 135.03156 , 129.84628 ,           
   150.3206  , 156.24223 , 190.24785 , 114.009544, 132.7888  ,           
   224.1297  , 105.6914  , 138.42787 , 213.81845 , 103.33176 ,           
   202.3564  , 161.14067 , 187.62227 , 176.96579 , 211.05275 ,           
   113.52303 ], dtype=float32)]                                         
Yes batch_normalization_12                                                   
[array([ 0.01501185,  0.29032254,  0.03455493,  0.07597651, -0.07218402,
   -0.16735457,  0.03332845, -0.00645222,  0.02006806, -0.03635798, 
    0.07917645, -0.05444119, -0.16789341, -0.086788  ,  0.12365555, 
   -0.19256745,  0.03382557,  0.04357202, -0.07722911,  0.03080065, 
   -0.13690165,  0.02521633,  0.04587894, -0.025393  , -0.1004908 , 
   -0.04962835,  0.09113129,  0.00759747,  0.2795353 , -0.23438764, 
   -0.1480339 , -0.02406131], dtype=float32), array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, 
   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, 
   nan, nan, nan, nan, nan, nan], dtype=float32),