📘

神经网络变得简单（第 73 部分）：价格走势预测 AutoBot·进阶篇

第 2/3 篇

本章目录

用 OpenCL 给转置层跑 GPU 加速
用OpenCL搭一个行情编码器的层栈
解码器前的多层注意力堆叠
堆叠转置与多头注意力构建特征交叉
初始化与退出时的模型装载校验
训练循环里的状态差分与批次截断

「用 OpenCL 给转置层跑 GPU 加速」

在 MT5 的神经网络封装里，CNeuronTransposeOCL 把矩阵转置丢给 GPU 内核，避免 CPU 端逐元素搬数据。Init 里先调基类 CNeuronBaseOCL::Init，传入 count * window 作为总元素数，再把 iWindow 与 iCount 存为成员，这两个值直接决定后面内核的全局工作维度。 Transpose 内核只用两行 id 取数：r = get_global_id(0)、c = get_global_id(1)，输出写回位置是 c * rows + r，等价于把输入 r * cols + c 的行列互换。feedForward 中 global_work_size 设为 {iCount, iWindow}，而反向的 calcInputGradients 把维度翻成 {iWindow, iCount}，转置的梯度本身也是一次转置。实盘接这套结构时，外汇与贵金属行情的高波动可能让批量矩阵维度频繁变化，建议在 MT5 策略测试器里先打印 iCount、iWindow 确认维度匹配，否则 OpenCL.Execute 会直接返回错误码并 printf 出 CL_ERROR_DESCRIPTION。

MQL5 / C++

class="type">uint window,
 ENUM_OPTIMIZATION optimization_type,
 class="type">uint batch)
{
 if(!CNeuronBaseOCL::Init(numOutputs, myIndex, open_cl, count * window,
                      optimization_type, batch))
    class="kw">return false;
class=class="str">"cmt">//---
 iWindow = window;
 iCount = count;
class=class="str">"cmt">//---
 class="kw">return true;
}
__kernel class="type">void Transpose(__global class="type">float *matrix_in,    class=class="str">"cmt">///<[in] Input matrix
                        __global class="type">float *matrix_out   class=class="str">"cmt">///<[out] Output matrix
                       )
{
 const class="type">int r = get_global_id(class="num">0);
 const class="type">int c = get_global_id(class="num">1);
 const class="type">int rows = get_global_size(class="num">0);
 const class="type">int cols = get_global_size(class="num">1);
class=class="str">"cmt">//---
 matrix_out[c * rows + r] = matrix_in[r * cols + c];
}
class="type">bool CNeuronTransposeOCL::feedForward(CNeuronBaseOCL *NeuronOCL)
{
 if(!NeuronOCL)
    class="kw">return false;
class=class="str">"cmt">//---
 class="type">uint global_work_offset[class="num">2] = {class="num">0, class="num">0};
 class="type">uint global_work_size[class="num">2] = {iCount, iWindow};
 if(!OpenCL.SetArgumentBuffer(def_k_Transpose, def_k_tr_matrix_in, NeuronOCL.getOutputIndex()))
    class="kw">return false;
 if(!OpenCL.SetArgumentBuffer(def_k_Transpose, def_k_tr_matrix_out, Output.GetIndex()))
    class="kw">return false;
 if(!OpenCL.Execute(def_k_Transpose, class="num">2, global_work_offset, global_work_size))
   {
     class="type">class="kw">string error;
     CLGetInfoString(OpenCL.GetContext(), CL_ERROR_DESCRIPTION, error);
     printf("Error of execution kernel Transpose: %d -> %s", GetLastError(), error);
     class="kw">return false;
   }
class=class="str">"cmt">//---
 class="kw">return true;
}
class="type">bool CNeuronTransposeOCL::calcInputGradients(CNeuronBaseOCL *NeuronOCL)
{
 if(!NeuronOCL)
    class="kw">return false;
class=class="str">"cmt">//---
 class="type">uint global_work_offset[class="num">2] = {class="num">0, class="num">0};
 class="type">uint global_work_size[class="num">2] = {iWindow, iCount};
 if(!OpenCL.SetArgumentBuffer(def_k_Transpose, def_k_tr_matrix_out, NeuronOCL.getGradientIndex()))
    class="kw">return false;
 if(!OpenCL.SetArgumentBuffer(def_k_Transpose, def_k_tr_matrix_in, Gradient.GetIndex()))
    class="kw">return false;
 if(!OpenCL.Execute(def_k_Transpose, class="num">2, global_work_offset, global_work_size))
   {
     class="type">class="kw">string error;

用OpenCL搭一个行情编码器的层栈

在MT5里跑价格行为模型，想把历史K线送进GPU做特征提取，得先按层把网络结构描述清楚。下面这段CreateTrajNetDescriptions函数就是干这事：先清掉旧描述，再从输入层一路堆到转置层，每一层都new一个CLayerDescription并塞进CArrayObj容器。输入层直接用defNeuronBaseOCL，节点数等于HistoryBars乘BarDescr，激活函数设None，优化器用ADAM。紧接着接BatchNorm层，batch尺寸取MathMax(1000, GPTBars)——也就是说样本数低于1000时强制按1000跑，显存占用和收敛稳定性得自己权衡。 Embedding层把prev_count长度的窗口压成GPTBars个token，输出维度交给EmbeddingSize。后面PEOCL做位置编码，MLMHAttentionOCL才是核心：step=4、window_out=16、layers=1，多头注意力的头数隐含在窗口拆分里。最后TransposeOCL翻一下张量维度，方便下游解码。这套层栈在外汇和贵金属行情上属于高波动数据建模，过拟合概率不低，实盘前务必用历史分笔回测验证。代码里任何一层Add失败就delete并return false，MT5终端会打出具体错因，直接照着改参数即可。

MQL5 / C++

class="type">bool CreateTrajNetDescriptions(CArrayObj *autobot)
  {
class=class="str">"cmt">//---
   CLayerDescription *descr;
class=class="str">"cmt">//---
   if(!autobot)
     {
      autobot = new CArrayObj();
      if(!autobot)
         class="kw">return false;
     }
class=class="str">"cmt">//--- Encoder
   autobot.Clear();
class=class="str">"cmt">//--- Input layer
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronBaseOCL;
   class="type">int prev_count = descr.count = (HistoryBars * BarDescr);
   descr.activation = None;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">1
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronBatchNormOCL;
   descr.count = prev_count;
   descr.batch = MathMax(class="num">1000,GPTBars);
   descr.activation = None;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">2
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronEmbeddingOCL;
     {
      class="type">int temp[] = {prev_count};
      ArrayCopy(descr.windows, temp);
     }
   prev_count = descr.count = GPTBars;
   class="type">int prev_wout = descr.window_out = EmbeddingSize;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">3
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronPEOCL;
   descr.count = prev_count;
   descr.window = prev_wout;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">4
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronMLMHAttentionOCL;
   descr.count = prev_count;
   descr.window = prev_wout;
   descr.step = class="num">4;
   descr.window_out = class="num">16;
   descr.layers = class="num">1;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">5
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronTransposeOCL;
   descr.count = prev_count;
   descr.window = prev_wout;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;

◍ 解码器前的多层注意力堆叠

这段构建逻辑里，从第 6 层到第 12 层连续往 autobot 里灌描述符，核心是用多头注意力（MLMHAttentionOCL）配合转置层（TransposeOCL）反复做维度置换。第 6、8、10 层结构几乎一致：count 取 prev_wout、window 取 prev_count，step 固定 4、window_out 固定 16、layers 为 1，优化器全走 ADAM。第 7 与第 9 层是纯转置，不挂优化器，只把 count/window 对调，给注意力层制造交叉视角的输入。这种对称穿插在 GPU 上跑 OCL 内核时，显存带宽占用倾向呈现周期性抖动。第 11 层突然切到 Concatenate，把 PrecoderBars * EmbeddingSize 作为 count，window 拉到 prev_count * prev_wout，step 等于 EmbeddingSize，激活用 LReLU。第 12 层才正式把 prev_count 重置为 PrecoderBars、prev_wout 重置为 EmbeddingSize，意味着解码器入口的张量形状在这里被锚定。任何一层 new 失败或 Add 返回否，都立刻 delete 并 return false，调用方拿不到半残网络。开 MT5 把 step 从 4 改成 2，可能看到推理延迟下降但过拟合概率上升，贵金属品种上这种敏感度高，属典型高风险调参。

MQL5 / C++

  class="kw">return false;
    }
class=class="str">"cmt">//--- layer class="num">6
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronMLMHAttentionOCL;
   descr.count = prev_wout;
   descr.window = prev_count;
   descr.step = class="num">4;
   descr.window_out = class="num">16;
   descr.layers = class="num">1;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">7
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronTransposeOCL;
   descr.count = prev_wout;
   descr.window = prev_count;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">8
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronMLMHAttentionOCL;
   descr.count = prev_count;
   descr.window = prev_wout;
   descr.step = class="num">4;
   descr.window_out = class="num">16;
   descr.layers = class="num">1;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">9
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronTransposeOCL;
   descr.count = prev_count;
   descr.window = prev_wout;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">10
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronMLMHAttentionOCL;
   descr.count = prev_wout;
   descr.window = prev_count;
   descr.step = class="num">4;
   descr.window_out = class="num">16;
   descr.layers = class="num">1;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- Decoder
class=class="str">"cmt">//--- layer class="num">11
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronConcatenate;
   descr.count = PrecoderBars * EmbeddingSize;
   descr.window = prev_count * prev_wout;
   descr.step = EmbeddingSize;
   descr.activation = LReLU;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">12
   if(!(descr = new CLayerDescription()))
      class="kw">return false;
   descr.type = defNeuronMLMHAttentionOCL;
   prev_count = descr.count = PrecoderBars;
   prev_wout = descr.window = EmbeddingSize;
   descr.step = class="num">4;
   descr.window_out = class="num">16;
   descr.layers = class="num">1;
   descr.optimization = ADAM;
   if(!autobot.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return false;
     }

「堆叠转置与多头注意力构建特征交叉」

这段网络组装代码连续注册了第 13 到 19 层，核心思路是用转置层调换特征与窗口维度，再交替塞入多头注意力层做跨维交互。外汇与贵金属行情里，这种结构可能更擅长捕捉不同周期粒度间的隐含关联，但高杠杆下模型过拟合风险同样偏高。第 14 层用 defNeuronMLMHAttentionOCL，step 写死为 4，window_out 固定 16，layers 为 1，优化器选 ADAM；第 16 层复用同型注意力但 count 与 window 参数对调，形成对称交叉。第 15 层是 Concatenate，把前层输出按 prev_count * prev_wout 拼接，step 指向 EmbeddingSize，激活用 LReLU；第 19 层收口用 defNeuronBaseOCL，节点数等于 PrecoderBars * 3，激活 None。每一层都先 new CLayerDescription，若 Add 失败立即 delete 并返回 false，这种写法能在 MT5 实盘加载时避免内存泄漏。打开你的 EA 源码把这段粘进网络初始化函数，改一下 PrecoderBars 看推理延迟变化。

MQL5 / C++

class=class="str">"cmt">//--- layer class="num">13
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronTransposeOCL;
  descr.count = prev_count;
  descr.window = prev_wout;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">14
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronMLMHAttentionOCL;
  descr.count = prev_wout;
  descr.window = prev_count;
  descr.step = class="num">4;
  descr.window_out = class="num">16;
  descr.layers = class="num">1;
  descr.optimization = ADAM;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">15
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronConcatenate;
  descr.count = prev_count * prev_wout;
  descr.window = descr.count;
  descr.step = EmbeddingSize;
  descr.activation = LReLU;
  descr.optimization = ADAM;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">16
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronMLMHAttentionOCL;
  descr.count = prev_count;
  descr.window = prev_wout;
  descr.step = class="num">4;
  descr.window_out = class="num">16;
  descr.layers = class="num">1;
  descr.optimization = ADAM;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">17
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronTransposeOCL;
  descr.count = prev_count;
  descr.window = prev_wout;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">18
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronMLMHAttentionOCL;
  descr.count = prev_wout;
  descr.window = prev_count;
  descr.step = class="num">4;
  descr.window_out = class="num">16;
  descr.layers = class="num">1;
  descr.optimization = ADAM;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//--- layer class="num">19
  if(!(descr = new CLayerDescription()))
      class="kw">return false;
  descr.type = defNeuronBaseOCL;
  descr.count = PrecoderBars * class="num">3;
  descr.activation = None;
  descr.optimization = ADAM;
  if(!autobot.Add(descr))
     {
       class="kw">delete descr;
       class="kw">return false;
     }
class=class="str">"cmt">//---
  class="kw">return true;
  }
CNet                Autobot;
class="type">int OnInit()
  {

初始化与退出时的模型装载校验

EA 在 OnInit 里先调 ResetLastError 清掉错误码，再尝试 LoadTotalBase 读取历史学习数据；若返回 false 就打印错误码并直接 INIT_FAILED，这一步决定了后续神经网络能否拿到训练底表。接着用 Autobot.Load 装载 Traj.nnw 权重文件，三个 temp 浮点占位、dtStudied 传入学习时间戳。文件不存在时走 CreateTrajNetDescriptions 新建网络描述并 Autobot.Create 初始化，任何一步失败都 delete 临时对象并回 INIT_FAILED。装载后必须做维度对账：getResults 输出总数应等于 PrecoderBars*3，第一层输出应等于 HistoryBars*BarDescr，否则打印不匹配数量并失败。这组硬校验能避免你改了输入窗口却忘了重训模型导致的静默错乱。 OpenCL 上下文从 Autobot.GetOpenCL 取得，Ones 与 Gradient 两个 buffer 分别用 1 和 0 初始化并挂到 OpenCL；State 按 HistoryBars*BarDescr 清零。最后 EventChartCustom 发 Init 事件通知图表，返回 INIT_SUCCEEDED。 OnDeinit 里只在非 INITFAILED、非 RECOMPILE 时调 Autobot.Save 落盘 Traj.nnw，并 delete Result 与 OpenCL。注意：外汇与贵金属杠杆交易高风险，模型装载失败可能让 EA 完全不交易，需在 MT5 日志里确认错误码后再排查。

MQL5 / C++

  ResetLastError();
  if(!LoadTotalBase())
    {
      PrintFormat("Error of load study data: %d", GetLastError());
      class="kw">return INIT_FAILED;
    }
class=class="str">"cmt">//--- load models
  class="type">float temp;
  if(!Autobot.Load(FileName + "Traj.nnw", temp, temp, temp, dtStudied, true))
    {
      Print("Init new models");
      CArrayObj *autobot = new CArrayObj();
      if(!CreateTrajNetDescriptions(autobot))
        {
          class="kw">delete autobot;
          class="kw">return INIT_FAILED;
        }
      if(!Autobot.Create(autobot))
        {
          class="kw">delete autobot;
          class="kw">return INIT_FAILED;
        }
      class="kw">delete autobot;
      class=class="str">"cmt">//---
    }
  Autobot.getResults(Result);
  if(Result.Total() != PrecoderBars * class="num">3)
    {
      PrintFormat("The scope of the Autobot does not match the precoder bars(%d <> %d)",
PrecoderBars * class="num">3, Result.Total());
      class="kw">return INIT_FAILED;
    }
class=class="str">"cmt">//---
  Autobot.GetLayerOutput(class="num">0, Result);
  if(Result.Total() != (HistoryBars * BarDescr))
    {
      PrintFormat("Input size of Autobot doesn&class="macro">#x27;t match state description(%d <> %d)",
Result.Total(), (HistoryBars * BarDescr));
      class="kw">return INIT_FAILED;
    }
  OpenCL = Autobot.GetOpenCL();
  if(!Ones.BufferInit(EmbeddingSize, class="num">1) ||
      !Gradient.BufferInit(EmbeddingSize, class="num">0) ||
      !Ones.BufferCreate(OpenCL) ||
      !Gradient.BufferCreate(OpenCL))
    {
      PrintFormat("Error of create buffers: %d", GetLastError());
      class="kw">return INIT_FAILED;
    }
  State.BufferInit(HistoryBars * BarDescr, class="num">0);
  if(!EventChartCustom(ChartID(), class="num">1, class="num">0, class="num">0, "Init"))
    {
      PrintFormat("Error of create study event: %d", GetLastError());
      class="kw">return INIT_FAILED;
    }
class=class="str">"cmt">//---
  class="kw">return(INIT_SUCCEEDED);
  }
class="type">void OnDeinit(const class="type">int reason)
  {
class=class="str">"cmt">//---
  if(!(reason == REASON_INITFAILED || reason == REASON_RECOMPILE))
    Autobot.Save(FileName + "Traj.nnw", class="num">0, class="num">0, class="num">0, TimeCurrent(), true);
  class="kw">delete Result;
  class="kw">delete OpenCL;
  }
class="type">void Train(class="type">void)
  {
class=class="str">"cmt">//---
  vector<class="type">float> probability = GetProbTrajectories(Buffer, class="num">0.9);
  vector<class="type">float> result, target, inp;
  matrix<class="type">float> targets;

◍ 训练循环里的状态差分与批次截断

这段逻辑跑在蒙特卡洛式采样的主循环里，外层以 Iterations 为上限，同时受 IsStopped() 与内部 Stop 标志约束，避免在 MT5 策略测试器点停止后还空转烧 CPU。每次迭代先用 SampleTrajectory(probability) 抽一条轨迹，再用双重 MathRand() 平方归一化挑起始 state，把随机性压到 [0,1) 且偏向低端，样本更集中在近期区段。若算出的 state 为负就 iter-- 并重抽，不浪费计数。 batch 定为 GPTBars+50，end 用 MathMin 卡在 Buffer[tr].Total - PrecoderBars 之前，确保后面塞 PrecoderBars 根预处理棒线时不越界。delta 矩阵按 (end-state-1) 行建，每行存状态向量；行号 row>0 时写的是「上一行 delta 减当前 inp」，即一阶差分而非原始状态——网络吃的是变化量。喂网前 Autobot.Clear() 清状态，feedForward 带 Ones 偏置矩阵且 false 表示不训练只前向；任一步返回失败就 PrintFormat 打函数行号、置 Stop=true 并 break。targets 按 PrecoderBars×3 建，若 size>BarDescr 会把整段棒线 reshape 后再裁到末行 3 列，只取最新聚合特征喂给预测头。

MQL5 / C++

  matrix<class="type">float> delta;
  STE = vector<class="type">float>::Zeros(PrecoderBars * class="num">3);
  class="type">int std_count = class="num">0;
  class="type">int batch = GPTBars + class="num">50;
  class="type">bool Stop = false;
  class="type">uint ticks = GetTickCount();
  class="type">ulong size = HistoryBars * BarDescr;
  for(class="type">int iter = class="num">0; (iter < Iterations && !IsStopped() && !Stop); iter ++)
    {
      class="type">int tr = SampleTrajectory(probability);
      class="type">int state = (class="type">int)((MathRand() * MathRand() / MathPow(class="num">32767, class="num">2)) *
(Buffer[tr].Total - class="num">3 - PrecoderBars - batch));
      if(state < class="num">0)
        {
         iter--;
         class="kw">continue;
        }
      class="type">int end = MathMin(state + batch, Buffer[tr].Total - PrecoderBars);
      Autobot.Clear();
      delta = matrix<class="type">float>::Zeros(end - state - class="num">1, Buffer[tr].States[state].state.Size());
      for(class="type">int i = state; i < end; i++)
        {
         inp.Assign(Buffer[tr].States[i].state);
         State.AssignArray(inp);
         if(i < (end - class="num">1))
           delta.Row(inp, row);
         if(row > class="num">0)
           delta.Row(delta.Row(row - class="num">1) - inp, row - class="num">1);
         if(!Autobot.feedForward((CBufferFloat*)GetPointer(State), class="num">1, false,
(CBufferFloat*)GetPointer(Ones)))
           {
            PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
            Stop = true;
            break;
           }
         targets = matrix<class="type">float>::Zeros(PrecoderBars, class="num">3);
         for(class="type">int t = class="num">0; t < PrecoderBars; t++)
           {
            target.Assign(Buffer[tr].States[i + class="num">1 + t].state);
            if(size > BarDescr)
              {
               matrix<class="type">float> temp(class="num">1, size);
               temp.Row(target, class="num">0);
               temp.Reshape(size / BarDescr, BarDescr);
               temp.Resize(size / BarDescr, class="num">3);
               target = temp.Row(temp.Rows() - class="num">1);
              }
            targets.Row(target, t);
           }
         targets.Reshape(class="num">1, targets.Rows()*targets.Cols());
         target = targets.Row(class="num">0);