From d3bfd8d700ce48a4af45b5d76a0642f1c96bd761 Mon Sep 17 00:00:00 2001
From: Mark <mah54746929@qq.com>
Date: Fri, 20 Mar 2020 18:57:10 +0800
Subject: [PATCH] innit

---
 main.py |  61 +++++++++++++++++++++++++++++++++
 nn.py   | 103 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 164 insertions(+)
 create mode 100644 main.py
 create mode 100644 nn.py

diff --git a/main.py b/main.py
new file mode 100644
index 0000000..b359aa8
--- /dev/null
+++ b/main.py
@@ -0,0 +1,61 @@
+
+import numpy as np
+
+from nn import Tnn
+
+Num     = 30
+x0      = np.linspace(0,1,Num)
+y0      = np.linspace(0,1,Num)
+xx,yy   = np.meshgrid(x0,y0)
+#zz1      = np.sqrt(4**2.0 - (xx**2.0 + yy**2.0)) / 2.0
+#zz2      = np.sqrt(2**2.0 - (xx**2.0 + yy**2.0)) / np.sqrt(2)
+
+zz1     = xx*yy
+zz2     = xx/2.0 + yy / 2.0
+
+input = np.c_[xx.reshape((Num*Num,1)).copy(),yy.reshape((Num*Num,1)).copy()]
+output= np.c_[zz1.reshape((Num*Num,1)).copy(), zz2.reshape((Num*Num,1)).copy()]
+
+test_input = np.c_[np.random.random(10),np.random.random(10)]
+
+
+if __name__ == '__main__':
+    nt0 = Tnn(2,10,2)
+    n_iteration = 100
+    
+    lr_cal = 1.0
+    anneal = 0.999
+       
+    # start loop
+    for i0 in range(n_iteration):
+    
+        err_cal = 0;
+        for j0 in range(output.shape[0]):
+            err_cal += nt0.train(input[j0,:], output[j0,:], lr_cal)
+        
+        #print 'Error: ', err_cal / output.shape[0], ' lr: ', lr_cal
+        lr_cal *= anneal
+        
+        # validation
+        test_prd = np.array([])
+        
+        for k0 in range(test_input.shape[0]):
+            if(k0==0):
+                test_prd = np.append(test_prd, nt0.predict(test_input[k0,:]))
+            else:
+                test_prd = np.vstack([test_prd, nt0.predict(test_input[k0,:])])
+                
+        # ground true
+        test_gr  = np.c_[test_input[:,0].copy()*test_input[:,1].copy(), test_input[:,1].copy()/2.0 + test_input[:,0].copy()/2.0]
+        
+        err_val = 0;
+        for k0 in range(test_prd.shape[0]):
+            for m0 in range(output.shape[1]):
+                err_val += 0.5*(test_prd[k0,m0] -test_gr[k0,m0])**2.0
+                
+        print 'Error: ', err_cal / output.shape[0], ' lr: ', lr_cal, 'Valid_Error: ', err_val / test_input.shape[0]
+            
+        
+    
+    
+    
\ No newline at end of file
diff --git a/nn.py b/nn.py
new file mode 100644
index 0000000..b7211cc
--- /dev/null
+++ b/nn.py
@@ -0,0 +1,103 @@
+import numpy as np
+
+class Tnn:
+    """my python verison of neural network"""
+    
+    nips=2
+    nops=2
+    nhis=10
+    
+    weights_hidden = np.array([])
+    weights_output = np.array([])
+    
+    bias = np.array([])
+    
+    # value stored at each layer
+    value_ops = np.array([])
+    value_his = np.array([])
+    
+    def __init__(self, nips_0, nhis_0, nops_0):
+        self.nips = nips_0
+        self.nops = nops_0
+        self.nhis = nhis_0
+        
+        self.weights_hidden = np.random.random(nips_0*nhis_0)
+        self.weights_output = np.random.random(nhis_0*nops_0)
+               
+        self.bias = np.random.random(2)
+        
+        # value stored at each layer
+        self.value_ops = np.zeros(nops_0)
+        self.value_his = np.zeros(nhis_0)
+    
+    # activation function
+    def actf(self, x_cal):
+        return 1.0/(1.0+np.exp(-x_cal))
+        
+    # partial derivative of activate function
+    def pdact(self, x_cal):
+        return x_cal * (1.0 - x_cal)
+        
+    # partial derivative of error function
+    def pderr(self, x_cal, y_cal):
+        return x_cal - y_cal
+        
+    # error function
+    def err(self, x_cal, y_cal):
+        return 0.5*(x_cal-y_cal)*(x_cal-y_cal)
+        
+    # compute total error of the target to output
+    def toerr(self, data_gt0, data_prd0):
+        sum = np.sum(self.err(data_gt0, data_prd0))
+        return sum
+        
+    def fprop(self, data_input):
+
+        # for hidden layer    
+        for i0 in range(self.nhis):
+            sum = 0
+            for j0 in range(self.nips):
+                sum += data_input[j0]*self.weights_hidden[j0+i0*self.nips]
+                
+            self.value_his[i0] = self.actf(sum + self.bias[0])
+            
+        # for output layer
+        for i0 in range(self.nops):
+            sum = 0
+            for j0 in range(self.nhis):
+                sum += self.value_his[j0]*self.weights_output[j0+i0*self.nhis]
+                
+            self.value_ops[i0] = self.actf(sum + self.bias[1])
+        
+    def bprop(self, data_input, data_gt, lr):
+        
+        for i0 in range(self.nhis):
+            sum = 0
+            # update weights of output layer
+            for j0 in range(self.nops):
+                a_cal = self.pderr(self.value_ops[j0], data_gt[j0])
+                b_cal = self.pdact(self.value_ops[j0])
+                
+                sum += a_cal * b_cal * self.weights_output[j0*self.nhis+i0]
+            
+                self.weights_output[j0*self.nhis+i0] -= lr * a_cal * b_cal * self.value_his[i0]
+            
+            # update weights in input layer            
+            for j0 in range(self.nips):
+                self.weights_hidden[i0*self.nips + j0] -= lr * sum * self.pdact(self.value_his[i0]) * data_input[j0]
+
+    def predict(self, data_input_test):
+        self.fprop(data_input_test)
+        return self.value_ops
+    
+    def train(self, data_input0, data_gt0, lr0):
+        self.fprop(data_input0)
+        self.bprop(data_input0, data_gt0, lr0)
+        return self.toerr(data_gt0, self.value_ops)
+        
+    def save_weights(self, name_s):
+        np.savetxt(name_s, np.c_[self.weights_hidden, self.weights_output])
+        
+    
+        
+    
\ No newline at end of file