对导数（梯度）传播的理解（上）

      

     导数的链式法则的理解，我参考了一篇 文章  Hacker's guide to Neural Networks

  

    

 var x = -2, y = 5,z=-4,h = 0.1;// whenever value of notation h was changeed, but the following derivateive value was not changed.
     var i=1;

      while (i<=10)
      {
        var out0=forwardCircuit(x, y, z);
        console.log("第"+i+"次输出为："+out0);
        var step_size=h;
        var der=derivative(x,y,z)
        x += step_size *  der.x_derivative;//
        y += step_size *  der.y_derivative;
        i++;
      }

     function derivative(x,y,z){
         var q=forwardAddGate(x,y);
         var  derivative_f_wrt_q = MultiplyGateDerivative(q,z,h).x_derivative;//(x+y)z 函数对 x+y函数的导数=z
         var derivative_q_wrt_x = AddGateDerivative(x,y,h).x_derivative;// x+y 函数对x 的导数为1
         var derivative_q_wrt_y = AddGateDerivative(x,y,h).y_derivative;// x+y 函数对y的导数为1
         //var x_derivative=MultiplyGateDerivative(x,y,h).x_derivative;
         //var y_derivative=MultiplyGateDerivative(x,y,h).y_derivative;
         var derivative_f_wrt_x = derivative_q_wrt_x * derivative_f_wrt_q; //(x+y)z 函数对 x  链式法则 =1xz
         var derivative_f_wrt_y = derivative_q_wrt_y * derivative_f_wrt_q; // (x+y)z 函数对 y  链式法则链式法则  =1xz
        console.log("乘法门的对x的斜率为"+MultiplyGateDerivative(q,z,h).x_derivative);
         console.log("乘法门的对y的斜率为"+derivative_f_wrt_q);
         console.log("加法门的对x的斜率为"+derivative_q_wrt_x);
        console.log("加法门的对y的斜率为"+derivative_q_wrt_y);
        console.log("走"+derivative_f_wrt_x);
        console.log("加法门的对y的斜率为"+derivative_q_wrt_y);
        return {x_derivative:derivative_f_wrt_x,y_derivative:derivative_f_wrt_y};
     }
      
      function MultiplyGateDerivative(x,y,h){//计算乘法的斜率             
                // compute derivative with respect to x
                var out = forwardMultiplyGate(x, y); // -6
                var xph = x + h; // 
                var out2 = forwardMultiplyGate(xph, y); // 
                var x_derivative = (out2 - out) / h; // 3.0
               // compute derivative with respect to y
                var yph = y + h; // 3.0001
                var out3 = forwardMultiplyGate(x, yph); // 
                var y_derivative = (out3 - out) / h; // -2.0
                return {x_derivative:x_derivative,y_derivative:y_derivative};
       }

      function AddGateDerivative(x,y,h){ // 计算加法的斜率      
                // compute derivative with respect to x
                var out = forwardAddGate(x, y); // -6
                var xph = x + h; // -1.9999
                var out2 = forwardAddGate(xph, y); // -5.9997
                var x_derivative = (out2 - out) / h; // 3.0
               // compute derivative with respect to y
                var yph = y + h; // 3.0001
                var out3 = forwardAddGate(x, yph); // -6.0002
                var y_derivative = (out3 - out) / h; // -2.0
                return {x_derivative:x_derivative,y_derivative:y_derivative};
       }

      function forwardMultiplyGate(x, y) { return x * y; };//乘法向前传播
      function forwardAddGate(a, b) { return a + b;};//加法向前传播
      function forwardCircuit(x,y,z) { //(x+y)z  
            var q = forwardAddGate(x, y);
            var f = forwardMultiplyGate(q, z);
                return f;
         };	

     function GradientCheck(){// 梯度检查
         var x_derivative = (forwardCircuit(x+h,y,z) - forwardCircuit(x,y,z)) / h; // -4
var y_derivative = (forwardCircuit(x,y+h,z) - forwardCircuit(x,y,z)) / h; // -4
var z_derivative = (forwardCircuit(x,y,z+h) - forwardCircuit(x,y,z)) / h; // 3
     }

就以上代码进行了优化

var x = -7, y = 6,z=-4,h = 0.1;// whenever value of notation h was changeed, but the following derivateive value was not changed.  
     var i=1;  
  
   while (i<=10)  
      {  
        var out0=forward(x, y, z);  
        console.log("第"+i+"次输出为："+out0);  
        var step_size=h;  
        var der=derivative(x,y,z)  
        x += step_size *  der.x_derivative;//  
        y += step_size *  der.y_derivative;  
        i++;  
      } 

     
  
     function derivative(x,y,z){
         var  der_q = DerMul(forwardAdd(x,y),z).x_derivative;//(x+y)z 函数对 x+y函数的导数=z 
         var  der_z= DerMul(forwardAdd(x,y),z).y_derivative;
         var der_x =1* der_q;//DerAdd(x,y).x_derivative;// x+y 函数对x 的导数为1  //(x+y)z 函数对 x  链式法则 =1xz
         var der_y =1* der_q;// DerAdd(x,y).y_derivative;// x+y 函数对y的导数为1  // (x+y)z 函数对 y  链式法则链式法则  =1xz  
        return {x_derivative:der_x,y_derivative:der_y};  
     }  
        
      

       function DerMul(x,y){//计算乘法的斜率              
                return {x_derivative:y,y_derivative:x};  
       }  


       function DerAdd(x,y){ // 计算加法的斜率                      
                return {x_derivative:1,y_derivative:1};  
       }  

      function forwardMul(x, y) { return x * y; };//乘法向前传播  
      function forwardAdd(a, b) { return a + b;};//加法向前传播  
      function forward(x,y,z) { //(x+y)z    
            var q = forwardAdd(x, y);  
            var f = forwardMul(q, z);  
                return f;  
         };   
  
     function GradientCheck(){// 梯度检查  
             var x_derivative = (forward(x+h,y,z) - forward(x,y,z)) / h; // -4  
             var y_derivative = (forward(x,y+h,z) - forward(x,y,z)) / h; // -4  
            var z_derivative = (forward(x,y,z+h) - forward(x,y,z)) / h; // 3  
     }