Question

从日语翻译成英语。（因为日本知之甚少）

我正在使用DL4j来预测股价。无论反复试验多少，学习都不会进行。是什么原因？

由于将要读取的数据进行了归一化，所以我认为每个品牌都没有太大区别。

附上图片后，无论您学习多少小时，它都将保持不变。

[输入] 0股票开始价格
1股票价格高
2股票价格低
3股票价格收盘
4唯一值A
5唯一值B
Unique valueA和Unique valueB是是我设计的算法的计算结果。它代表了当日的特征，并与未来股价的涨跌有着明确的联系。

[输出] 目前，我们预计up``silent``down将在10天内产生三项输出。

即使您尝试使用两个输出up``other进行此操作，学习也不会继续进行。

预测股价在10天内上涨或下跌10％以上。

如前所述，输入[4]和输入[5]与股价飙升有关，并且可以预测。

我进行机器学习以消除误报。

【数据集】有巨大的数据集。这是一个例子。时间步长是100。

===========INPUT===================
[[[    1.0357,    1.0513,    1.0147  ...    0.9730    0.9695,    0.9734], 
  [    1.0388,    1.0341,    1.0022  ...    0.9656    0.9734,    0.9734], 
  [    1.0427,    1.0513,    1.0201  ...    0.9734    0.9734,    0.9738], 
  [    1.0318,    1.0287,    0.9855  ...    0.9656    0.9664,    0.9734], 
  [    1.6755,    1.2601,    0.3456  ...         0    0.7736,    6.7037], 
  [    0.3751,    0.3616,    1.1761  ...    0.2864    0.5629,    0.4131]], 

 [[    1.0821,    1.0801,    1.0601  ...    1.3061    1.3001,    1.2961], 
  [    1.1001,    1.0901,    1.0401  ...    1.2981    1.2821,    1.3381], 
  [    1.1001,    1.0901,    1.0601  ...    1.3101    1.3001,    1.3381], 
  [    1.0801,    1.0621,    1.0241  ...    1.2601    1.2821,    1.2961], 
  [    0.0688,    0.9349,    0.0838  ...    1.6641         0,    0.3009], 
  [    1.3479,    0.0673,         0  ...    0.6315    0.9210,   58.8286]], 

 [[    1.0900,    1.0708,    1.0708  ...    1.0900    1.0836,    1.0772], 
  [    1.0772,    1.0580,    1.1157  ...    1.0836    1.0708,    1.0516], 
  [    1.0964,    1.0708,    1.1157  ...    1.0900    1.0836,    1.0772], 
  [    1.0772,    1.0451,    1.0708  ...    1.0772    1.0708,    1.0516], 
  [    0.2267,    0.9888,         0  ...    0.5150    2.3653,         0], 
  [    1.7519,    0.6446,    1.7559  ...    2.0206    0.2649,    9.3929]], 

  ..., 

 [[    0.9925,    0.9925,    0.9953  ...    1.0109    1.0091,    1.0119], 
  [    0.9925,    0.9971,    0.9888  ...    1.0063    1.0119,    1.0082], 
  [    0.9971,    0.9980,    0.9953  ...    1.0119    1.0119,    1.0128], 
  [    0.9925,    0.9879,    0.9852  ...    1.0063    1.0073,    1.0063], 
  [    0.5251,    0.5053,    0.4411  ...    1.9785    0.9391,    7.1644], 
  [    0.9026,         0,    0.8412  ...    0.1732    0.1513,    0.0467]], 

 [[    0.7903,    0.7764,    0.7875  ...    1.0193    1.0165,    0.9216], 
  [    0.7764,    0.8015,    0.7819  ...    1.0221    1.0109,    0.9160], 
  [    0.7959,    0.8015,    0.7959  ...    1.0584    1.0556,    0.9355], 
  [    0.7764,    0.7764,    0.7819  ...    1.0193    1.0054,    0.8881], 
  [    0.7363,    0.3788,    0.3158  ...    0.4995    0.6567,    6.4159], 
  [    0.1783,    0.3246,    0.8313  ...    0.6535    0.3348,    0.1456]], 

 [[    0.8960,    0.9086,    0.9194  ...    0.9844    0.9553,    0.9875], 
  [    0.8935,    0.9169,    0.9068  ...    0.9680    0.9585,    1.0159], 
  [    0.8960,    0.9219,    0.9225  ...    0.9869    0.9831,    1.0229], 
  [    0.8821,    0.9023,    0.9017  ...    0.9642    0.9553,    0.9818], 
  [    4.5974,         0,    0.6352  ...    0.4522    0.6429,    0.7656], 
  [    0.3374,    0.3450,    1.0265  ...    1.5891    2.2172,    7.6396]]]
=================OUTPUT==================
[[[         0,         0,         0  ...         0         0,         0], 
  [         0,         0,         0  ...         0         0,    1.0000], 
  [         0,         0,         0  ...         0         0,         0]], 

 [[         0,         0,         0  ...         0         0,    1.0000], 
  [         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,    1.0000]], 

  ..., 

 [[         0,         0,         0  ...         0         0,         0], 
  [         0,         0,         0  ...         0         0,         0], 
  [         0,         0,         0  ...         0         0,    1.0000]], 

 [[         0,         0,         0  ...         0         0,         0], 
  [         0,         0,         0  ...         0         0,    1.0000], 
  [         0,         0,         0  ...         0         0,         0]], 

 [[         0,         0,         0  ...         0         0,    1.0000], 
  [         0,         0,         0  ...         0         0,         0], 
  [         0,         0,         0  ...         0         0,         0]]]
===========INPUT MASK===================
[[    1.0000,    1.0000,    1.0000  ...    1.0000    1.0000,    1.0000], 
 [    1.0000,    1.0000,    1.0000  ...    1.0000    1.0000,    1.0000], 
 [    1.0000,    1.0000,    1.0000  ...    1.0000    1.0000,    1.0000], 
  ..., 
 [    1.0000,    1.0000,    1.0000  ...    1.0000    1.0000,    1.0000], 
 [    1.0000,    1.0000,    1.0000  ...    1.0000    1.0000,    1.0000], 
 [    1.0000,    1.0000,    1.0000  ...    1.0000    1.0000,    1.0000]]
===========OUTPUT MASK===================
[[         0,         0,         0  ...         0         0,    1.0000], 
 [         0,         0,         0  ...         0         0,    1.0000], 
 [         0,         0,         0  ...         0         0,    1.0000], 
  ..., 
 [         0,         0,         0  ...         0         0,    1.0000], 
 [         0,         0,         0  ...         0         0,    1.0000], 
 [         0,         0,         0  ...         0         0,    1.0000]]

【神经网络】

NumInputs = 6
NumOutputs = 3
NumLstmLayers = 256
val conf = NeuralNetConfiguration.Builder()
                .seed(19920528)
                .weightInit(WeightInit.XAVIER)
                .miniBatch(true)
                .updater(Adam())
                .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)

                .list()
                .layer(LSTM.Builder()
                        .nIn(NumInputs)
                        .nOut(NumLstmLayers)
                        .activation(Activation.TANH)
                        .build()
                )
                .layer(LSTM.Builder()
                        .nIn(NumLstmLayers)
                        .nOut(NumLstmLayers)
                        .activation(Activation.TANH)
                        .build()
                )
                .layer(RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT)
                        .nIn(NumLstmLayers)
                        .nOut(NumOutputs)
                        .activation(Activation.SOFTMAX)
                        .build()
                )

//               .backpropType(BackpropType.TruncatedBPTT)
//                .tBPTTForwardLength(5)
//                .tBPTTBackwardLength(5)
                .build()

        val nn = MultiLayerNetwork(conf)
        nn.init()

【其他信息】小批量= 32 时代= 10

RNN根本无法学习

0 个答案: