Added XNOR and removed some mistakes

LogisticRegression/GradientAscent-Log-Lhood/XOR/main.py

@@ -12,7 +12,7 @@ def sigmoid(z):
 def loss(X,theta,Y) :
     scores = np.dot(X, theta.T)
     ll = np.sum( Y.reshape(4,1)*scores - np.log(1 + np.exp(scores)) )
-    return ll
+    return -ll
 
 
 def maxima(X,Y,theta,a,it):
@@ -38,31 +38,32 @@ def maxima(X,Y,theta,a,it):
 print(X)
 
 
-a=1
-it=500000
+a=9
+it=800000
 theta=np.array([[0,0,0,0]])
 print(theta.shape,X.shape,Y.shape)
 theta=maxima(X,Y,theta,a,it)
 print(theta)
 
-xx, yy =np.meshgrid(np.arange(-1,2,0.5), np.arange(-1,2,0.5))
+xx, yy =np.meshgrid(np.arange(-0.2,1.2,0.01), np.arange(-0.2,1.2,0.01))
 
 
-z=((-theta[0,0]*xx-theta[0,1]*yy-theta[0,3])/theta[0,2])
-
+z=+theta[0,0]*xx.ravel()+theta[0,1]*yy.ravel()+theta[0,3]+theta[0,2]*(np.multiply(xx,yy).ravel())
 
 
 ###########
 
 
 fig = plt.figure()
-ax = Axes3D(fig)
-ax.scatter(X[0:2,0],X[0:2,1],X[0:2,2],c='r',marker='x')
 
-ax.scatter(X[2:,0],X[2:,1],X[2:,2],c='b',marker='o')
+plt.scatter(xx.ravel(),yy.ravel(),c=z)
+
+plt.scatter(X[0:2,0],X[0:2,1],c='r',marker='x',s=40)
+
+plt.scatter(X[2:,0],X[2:,1],c='b',marker='o',s=40)
+
+
 
-ax.plot_surface(xx,yy,z,alpha=0.9,cmap=cm.summer)
-ax.set_zlim(0,1.2)
 
 plt.show()
 

XOR_and_XNOR_Neural_Network/XNOR_CrossEntropy.py

@@ -0,0 +1,90 @@
+import numpy as np
+import matplotlib.pyplot as plt
+np.random.seed(0)
+X= np.array([[0,0,1,1],
+            [0,1,0,1]])
+Y=np.array([[1,0,0,1]])
+
+
+
+b1=np.random.random()
+b2=np.random.random()
+
+def sigmoidgrad(A):
+    s=1/(1+np.exp(-A))
+    return s*(1-s)
+
+def sigmoid(A):
+    return 1/(1+np.exp(-A))
+
+theta1 = (np.random.random(( 2, 2)))
+theta2 = (np.random.random((2, 1)))
+
+def forward(X,theta1,theta2,b1,b2):
+    a1=X
+    z1=theta1.T.dot(a1)+b1
+    a2=sigmoid(z1)
+    z2=theta2.T.dot(a2)+b2
+    hyp=sigmoid(z2)
+    return a1,z1,a2,z2,hyp
+
+
+###Back Propagation###
+
+lr=3;
+mse=[]
+for i in range(1900):
+
+    a1, z1, a2, z2, hyp = forward(X, theta1, theta2,b1,b2)
+    dw2=a2.dot((hyp-Y).T)
+    dw1=np.multiply(theta2.dot(hyp-Y),sigmoidgrad(z1)).dot(a1.T)
+    db2=hyp-Y
+    db1=np.multiply(theta2.dot(hyp-Y),sigmoidgrad(z1))
+    #Gradient Checking code
+    #a1, z1, a2, z2, hyp1 = forward(X, theta1, theta2,b1+0.001,b2)
+    #a1, z1, a2, z2, hyp2 = forward(X, theta1, theta2,b1-0.001,b2)
+
+    #print(np.sum(db1),'--')
+    #print( np.sum((-(Y*(np.log(hyp1))+(1-Y)*(np.log(1-hyp1))))-(-(Y*(np.log(hyp2))+(1-Y)*(np.log(1-hyp2)))),axis=1)/0.002 )
+
+    theta1 -= lr*dw1
+    theta2 -= lr*dw2
+    b1 -= lr*np.sum(db1)
+    b2 -= lr*np.sum(db2)
+
+
+    a1, z1, a2, z2, hyp = forward(X, theta1, theta2,b1,b2)
+    if i%100 ==0:    
+        print((1/8)*np.sum((Y-hyp)**2))
+    mse.append([i,(1/8)*np.sum((Y-hyp)**2)])
+
+a,b,c,d,e = forward(X,theta1,theta2,b1,b2)
+print(e)
+
+mse=np.array(mse)
+plt.figure(figsize=[10,5])
+plt.subplot(121)
+plt.plot(mse[:,0],mse[:,1])
+plt.xlabel('Epochs')
+plt.ylabel('Mean Squared Error')
+plt.subplot(122)
+
+
+h = .02  # step size in the mesh
+x_min, x_max = X[0, :].min() - 0.2, X[0, :].max() + 0.2
+y_min, y_max = X[1, :].min() - 0.2, X[1, :].max() + 0.2
+xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
+
+
+xnew=np.concatenate((xx.ravel().reshape(1,4900),yy.ravel().reshape(1,4900)),axis=0)
+
+
+a,b,c,d,e = forward(xnew,theta1,theta2,b1,b2)
+plt.scatter(xnew[0,:],xnew[1,:],c=e[0,:])
+
+plt.scatter(X[0,0],X[1,0],c='w',edgecolors='w',marker='o',s=60)
+plt.scatter(X[0,1],X[1,1],c='w',edgecolors='w',marker='x',s=60)
+plt.scatter(X[0,2],X[1,2],c='w',edgecolors='w',marker='x',s=60)
+plt.scatter(X[0,3],X[1,3],c='w',edgecolors='w',marker='o',s=60)
+plt.show()
+