Simple XOR learning with keras

xor.py

from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.optimizers import SGD
import numpy as np 

X = np.array([[0,0],[0,1],[1,0],[1,1]])
y = np.array([[0],[1],[1],[0]])

model = Sequential()
model.add(Dense(8, input_dim=2))
model.add(Activation('tanh'))
model.add(Dense(1))
model.add(Activation('sigmoid'))

sgd = SGD(lr=0.1)
model.compile(loss='binary_crossentropy', optimizer=sgd)

model.fit(X, y, show_accuracy=True, batch_size=1, nb_epoch=1000)
print(model.predict_proba(X))
"""
[[ 0.0033028 ]
 [ 0.99581173]
 [ 0.99530098]
 [ 0.00564186]]
"""

Great simple example. I would get it to work with only two neurons in the dense layer by running for more epochs. I tested this and it gets to similar accuracy with 5000 epochs. With rounding, only ~2000 epochs are needed.

Try this code. Best result I was able to get from it is 22 epochs with batchsize = 4. So only 86 network evaluations !!!

from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.optimizers import SGD
from keras.callbacks import Callback
from keras.initializers import VarianceScaling 
import numpy as np 


lastEpoch = 0


class EarlyStoppingByLossVal(Callback):
    def __init__(self, monitor='val_loss', value=0.008, verbose=0):
        super(Callback, self).__init__()
        self.monitor = monitor
        self.value = value
        self.verbose = verbose
    def on_epoch_end(self, epoch, logs={}):
        global lastEpoch
        current = logs.get("loss")         
        if current != None and current < self.value:
            self.model.stop_training = True
            lastEpoch = epoch + 1


x = np.array([
    [0,0], [0,1],
    [1,0], [1,1]
])
y = np.array([
    [0], [1], 
    [1], [0]
])

model = Sequential()
model.add(Dense(8, 
                input_dim = 2, 
                use_bias = False, 
                kernel_initializer = VarianceScaling()))
model.add(Activation('tanh'))
model.add(Dense(1, 
                use_bias = False, 
                kernel_initializer = VarianceScaling()))
model.add(Activation('tanh'))
model.compile(loss = "mean_squared_error", 
              optimizer = SGD(lr = 0.6, 
                              momentum = 0.6))

model.fit(x, y, 
          verbose = 1, 
          batch_size = 4, 
          epochs = 10000, 
          callbacks = [
            EarlyStoppingByLossVal()
          ])

print(model.predict_proba(x))
print("Last epoch: " + str(lastEpoch))

how would i solve this xor problem:

n=300
x1 = np.random.rand(n,2) * (-1)
x2 = np.random.rand(n,2)
x2[:,1] *= (-1)

x3 = np.random.rand(n,2)
x3[:,0] *= (-1)
x4 = np.random.rand(n,2)
x = np.concatenate((x1, x2, x3, x4))
x = (x + 1 ) /2
y1 = np.ones(n)
y4 = np.ones(n)
y2 = np.zeros(n)
y3 = np.zeros(n)
y = np.concatenate((y1,y2,y3,y4))
print (x1[[1,2],:])
print (x2[[1,2],:])
print (x3[[1,2],:])
print (x4[[1,2],:])

import matplotlib.pyplot as plt

plt.plot(x1[:,0], x1[:,1], 'ro')
plt.plot(x2[:,0], x2[:,1], 'bo')
plt.plot(x3[:,0], x3[:,1], 'bo')
plt.plot(x4[:,0], x4[:,1], 'ro')

plt.show()

Do you know how to build a xor model (or other binary task) using simple recurrent layers? Does it has any sense to do that? Could you comment regarding that?

baj12 left us a good example. While working on it (I added some noise) I could not get loss values less than 0.22...
Any help to get better loss values?
PS: if you want to change from SOFTMAX to SIGMOID activation you should remove categorical from y.

n= 200
ruido = 3
fat = n*ruido/100
print("nivel de ruido",ruido,"%")

x1 = np.random.rand(n,2) * (-1)
plt.plot(x1[:,0], x1[:,1], 'ro')
x11 = np.random.rand(fat,2) * (-1)
plt.plot(x11[:,0], x11[:,1], 'bo')
x2 = np.random.rand(n,2)
x2[:,1] *= (-1)
plt.plot(x2[:,0], x2[:,1], 'bo')
x22 = np.random.rand(fat,2)
x22[:,1] *= (-1)
plt.plot(x22[:,0], x22[:,1], 'ro')
x3 = np.random.rand(n,2)
x3[:,0] *= (-1)
plt.plot(x3[:,0], x3[:,1], 'bo')
x33 = np.random.rand(fat,2)
x33[:,0] *= (-1)
plt.plot(x33[:,0], x33[:,1], 'ro')
x4 = np.random.rand(n,2)
plt.plot(x4[:,0], x4[:,1], 'ro')
x44 = np.random.rand(fat,2)
plt.plot(x44[:,0], x44[:,1], 'bo')
X = np.concatenate((x1,x11,x2,x22,x3,x33,x4,x44))
X = (X + 1 ) /2
y1 = np.ones(n)
y11= np.zeros(fat)
y4 = np.ones(n)
y44 = np.zeros(fat)
y2 = np.zeros(n)
y22 = np.ones(fat)
y3 = np.zeros(n)
y33 = np.ones(fat)
y = np.concatenate((y1,y11,y2,y22,y3,y33,y4,y44))

if you want to change from SOFTMAX to SIGMOID activation you should remove categorical from y.

y2 = np_utils.to_categorical(y)
#y = np_utils.to_categorical(y)
model = Sequential()
model.add(Dense(12, input_dim=X.shape[1], activation='tanh',kernel_initializer = VarianceScaling()))
model.add(Dense(2, init='uniform', activation='softmax', name="output"))
#model.add(Dense(2, init='uniform', activation='sigmoid', name="output"))
sgd = SGD(lr=0.01)
model.compile(loss='binary_crossentropy', optimizer=sgd)
model.summary()
model.fit(X, y2, batch_size=2, shuffle=True, nb_epoch=2000, verbose=1,callbacks =[EarlyStoppingByLossVal()])
#model.fit(X, y, batch_size=2, shuffle=True, nb_epoch=2000, verbose=1,callbacks =[EarlyStoppingByLossVal()])
plot_decision_boundary(lambda X:model.predict_classes(X))
print("Last epoch: " + str(lastEpoch))

sorry, forgot to include the libraries I used to run the code above

from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.optimizers import SGD
from keras.callbacks import Callback
from keras.initializers import VarianceScaling
import numpy as np
import matplotlib.pyplot as plt

lastEpoch = 0

class EarlyStoppingByLossVal(Callback):
def init(self, monitor='val_loss', value=0.02, verbose=0):
super(Callback, self).init()
self.monitor = monitor
self.value = value
self.verbose = verbose
def on_epoch_end(self, epoch, logs={}):
global lastEpoch
current = logs.get("loss")
if current != None and current < self.value:
self.model.stop_training = True
lastEpoch = epoch + 1

def plot_decision_boundary(pred_func):
# Set min and max values and give it some padding
x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5
y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5
h = 0.01
# Generate a grid of points with distance h between them
xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
# Predict the function value for the whole gid
Z = pred_func(np.c_[xx.ravel(), yy.ravel()])
Z = Z.reshape(xx.shape)
# Plot the contour and training examples
plt.contourf(xx, yy, Z, cmap=plt.cm.Spectral)
plt.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.Spectral)

I wanted to solve this with only two hidden units. So I used this code which worked fine (for most executions, depending on the random initial condition):

 from keras.models import Sequential
 from keras.layers.core import Dense, Dropout, Activation
 from keras.optimizers import SGD
 import numpy as np
 
 X = np.array([[0,0],[0,1],[1,0],[1,1]])
 y = np.array([[0],[1],[1],[0]])
 
 model = Sequential()
 model.add(Dense(2, input_dim=2))
 model.add(Activation('tanh'))
 model.add(Dense(1))
 model.add(Activation('sigmoid'))
 
 sgd = SGD(lr=0.1)
 model.compile(loss='mean_squared_error', optimizer=sgd)
 
 model.fit(X, y, batch_size=1, epochs=1000)
 print(model.predict_proba(X))

I think to solve it for any initial condition we need to have scattered inputs like @baj12 proposed. But I didn't test it.

I added some biases and random initialization as well, however I have no better result as consciencia

rndU = RandomUniform(minval=-1, maxval=1, seed=None)
model = Sequential()
model.add(Dense(9, activation='sigmoid', input_dim=2, use_bias = True, kernel_initializer=rndU, bias_initializer=rndU))
model.add(Dense(1, activation='sigmoid', use_bias = True, kernel_initializer=rndU, bias_initializer=rndU))

from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import SGD
from keras.initializers import RandomUniform
import numpy as np
x=np.array([[0.1,0.1,1],
[0.1,0.9,1],
[0.9,0.1,1],
[0.9,0.9,1]])
y=np.array([[0.1],[0.9],[0.9],[0.1]])
model= Sequential()
model.add(Dense(4,input_dim=3,activation="sigmoid",
bias_initializer=RandomUniform(minval=-1.0, maxval=1, seed=None)))
model.add(Dense(1,activation="sigmoid",bias_initializer=RandomUniform(minval=-1.0, maxval=1, seed=None)))
sgd=SGD(lr=0.01)
model.compile(loss='mean_squared_error',optimizer='sgd')
model.fit(x,y,epochs=5000,batch_size=1,verbose=1)

i am not geeting good result what i am doing wrong any idea

@gauravkr0071 replace
model.compile(loss='mean_squared_error',optimizer='sgd')
by this
model.compile(loss='mean_squared_error',optimizer=sgd)