

In [1]:

    
%pylab inline
from classy import *









    



Populating the interactive namespace from numpy and matplotlib
Version:  0.0.16

Iris



In [2]:

    
data=load_excel('data/iris.xls')
data_train,data_test=split(data,test_size=0.2)









    



iris.data 151 5
150 vectors of length 4
Feature names: 'petal length in cm', 'petal width in cm', 'sepal length in cm', 'sepal width in cm'
Target values given.
Target names: 'Iris-setosa', 'Iris-versicolor', 'Iris-virginica'
Mean:  [ 3.75866667  1.19866667  5.84333333  3.054     ]
Median:  [ 4.35  1.3   5.8   3.  ]
Stddev:  [ 1.75852918  0.76061262  0.82530129  0.43214658]
Original vector shape:  (150, 4)
Train vector shape:  (120, 4)
Test vector shape:  (30, 4)



In [3]:

    
len(data.targets),len(data_train.targets),len(data_test.targets),









    Out[3]:





(150, 120, 30)



In [4]:

    
C=Perceptron()



In [5]:

    
timeit(reset=True)
C.fit(data_train.vectors,data_train.targets)
print(("Training time: ",timeit()))









    



Time Reset
('Training time: ', '0.006078004837036133 seconds ')



In [6]:

    
print(("On Training Set:",C.percent_correct(data_train.vectors,data_train.targets)))
print(("On Test Set:",C.percent_correct(data_test.vectors,data_test.targets)))









    



('On Training Set:', 66.666666666666657)
('On Test Set:', 66.666666666666657)



In [7]:

    
C.coef_  # these are the weights









    Out[7]:





array([[ -8.4,  -3.4,   1.3,   5.3],
       [  1.1,  -6.1,   0.5, -10. ],
       [ 31.6,  27.6, -21.2, -17.5]])



In [8]:

    
C=BackProp(hidden_layer_sizes = [4],max_iter=10000,tol=1e-4)



In [9]:

    
timeit(reset=True)
C.fit(data_train.vectors,data_train.targets)
print(("Training time: ",timeit()))









    



Time Reset
('Training time: ', '0.6452150344848633 seconds ')



In [10]:

    
data_train.vectors.shape,data_train.targets.shape









    Out[10]:





((120, 4), (120,))



In [11]:

    
print(("On Training Set:",C.percent_correct(data_train.vectors,data_train.targets)))
print(("On Test Set:",C.percent_correct(data_test.vectors,data_test.targets)))









    



('On Training Set:', 98.333333333333329)
('On Test Set:', 96.666666666666671)



In [12]:

    
C.weights









    Out[12]:





[array([[  5.01417920e-01,   7.24123351e-13,   5.24105291e-01,
          -1.31535751e+00],
        [  1.18622033e+00,  -4.66957528e-07,  -9.15263078e-02,
          -1.51182807e+00],
        [  7.38937254e-01,  -1.57798905e-03,  -3.57895679e-01,
           8.74776410e-01],
        [ -4.73245427e-02,   3.31502216e-06,   1.27632460e-01,
           1.06797780e+00]]),
 array([[ -7.90372299e-01,   1.07819361e-01,   6.52942510e-01],
        [ -1.77117250e-09,   4.47442479e-06,  -1.40396531e-39],
        [  7.51372708e-01,   7.21944832e-01,  -2.25284488e-01],
        [  1.56579181e+00,   5.78388845e-01,  -1.54632369e+00]])]



In [13]:

    
W_inp_hid,W_hid_out=C.weights
print(W_inp_hid)
print("==")
print(W_hid_out)









    



[[  5.01417920e-01   7.24123351e-13   5.24105291e-01  -1.31535751e+00]
 [  1.18622033e+00  -4.66957528e-07  -9.15263078e-02  -1.51182807e+00]
 [  7.38937254e-01  -1.57798905e-03  -3.57895679e-01   8.74776410e-01]
 [ -4.73245427e-02   3.31502216e-06   1.27632460e-01   1.06797780e+00]]
==
[[ -7.90372299e-01   1.07819361e-01   6.52942510e-01]
 [ -1.77117250e-09   4.47442479e-06  -1.40396531e-39]
 [  7.51372708e-01   7.21944832e-01  -2.25284488e-01]
 [  1.56579181e+00   5.78388845e-01  -1.54632369e+00]]

XOR Problem - Perceptron



In [14]:

    
data=load_csv('data/xor.csv')
print() 
print((data.vectors))
print() 
print((data.targets))









    



4 vectors of length 2
Feature names: 'p1', 'p2'
Target values given.
Target names: '0', '1'
Mean:  [ 0.5  0.5]
Median:  [ 0.5  0.5]
Stddev:  [ 0.5  0.5]

[[ 0.  0.]
 [ 1.  0.]
 [ 0.  1.]
 [ 1.  1.]]

[0 1 1 0]



In [15]:

    
C=Perceptron()



In [16]:

    
C.fit(data.vectors,data.targets)



In [17]:

    
print((C.predict(data.vectors)))
print(("On Training Set:",C.percent_correct(data.vectors,data.targets)))









    



[0 0 0 0]
('On Training Set:', 50.0)



In [18]:

    
plot2D(data,classifier=C,axis_range=[-.55,1.5,-.5,1.5])

XOR Problem - Backprop



In [19]:

    
data.vectors









    Out[19]:





array([[ 0.,  0.],
       [ 1.,  0.],
       [ 0.,  1.],
       [ 1.,  1.]])



In [20]:

    
data.targets









    Out[20]:





array([0, 1, 1, 0])



In [21]:

    
C=BackProp(hidden_layer_sizes = [5],max_iter=10000,tol=1e-4)



In [22]:

    
C.fit(data.vectors,data.targets)



In [23]:

    
print((C.predict(data.vectors)))
print(("On Training Set:",C.percent_correct(data.vectors,data.targets)))









    



[0 1 1 0]
('On Training Set:', 100.0)



In [24]:

    
plot2D(data,classifier=C,axis_range=[-.55,1.5,-.5,1.5])



In [25]:

    
print((data.vectors))
print()
print((data.targets))









    



[[ 0.  0.]
 [ 1.  0.]
 [ 0.  1.]
 [ 1.  1.]]

[0 1 1 0]



In [26]:

    
C.output(data.vectors)









    Out[26]:





[array([[  1.01831113e+00,   1.90534269e+00,   0.00000000e+00,
           2.49285791e-01,   0.00000000e+00],
        [  1.31058231e+00,   2.31129715e-05,   0.00000000e+00,
           7.59467427e-04,   0.00000000e+00],
        [  0.00000000e+00,   0.00000000e+00,   0.00000000e+00,
           2.24620651e+00,   1.78930977e+00],
        [  0.00000000e+00,   0.00000000e+00,   0.00000000e+00,
           1.99768018e+00,   0.00000000e+00]]), array([[ 0.06029832],
        [ 0.92728785],
        [ 0.94054385],
        [ 0.06926277]])]



In [27]:

    
h,y=C.output(data.vectors)
print(h)
print() 
print((np.round(h)))
print()
print(y)









    



[[  1.01831113e+00   1.90534269e+00   0.00000000e+00   2.49285791e-01
    0.00000000e+00]
 [  1.31058231e+00   2.31129715e-05   0.00000000e+00   7.59467427e-04
    0.00000000e+00]
 [  0.00000000e+00   0.00000000e+00   0.00000000e+00   2.24620651e+00
    1.78930977e+00]
 [  0.00000000e+00   0.00000000e+00   0.00000000e+00   1.99768018e+00
    0.00000000e+00]]

[[ 1.  2.  0.  0.  0.]
 [ 1.  0.  0.  0.  0.]
 [ 0.  0.  0.  2.  2.]
 [ 0.  0.  0.  2.  0.]]

[[ 0.06029832]
 [ 0.92728785]
 [ 0.94054385]
 [ 0.06926277]]



In [29]:

    
C.weights









    Out[29]:





[array([[  2.92271179e-01,  -1.90531957e+00,  -2.83381318e-08,
          -2.48526323e-01,  -1.78979754e+00],
        [ -1.31122171e+00,  -1.90546166e+00,  -3.18856261e-19,
           1.99692072e+00,   1.78936911e+00]]), array([[  1.86464160e+00],
        [ -2.31509246e+00],
        [  3.52148889e-25],
        [ -1.35213790e+00],
        [  3.18297629e+00]])]



In [30]:

    
data.vectors.shape









    Out[30]:





(4, 2)

Curvy data



In [31]:

    
N=30
x1=randn(N)*.2
y1=randn(N)*.2

plot(x1,y1,'bo')

a=linspace(0,3*pi/2,N)
x2=cos(a)+randn(N)*.2
y2=sin(a)+randn(N)*.2

plot(x2,y2,'rs')









    Out[31]:





[<matplotlib.lines.Line2D at 0x1163b6ef0>]



In [32]:

    
vectors=vstack([hstack([atleast_2d(x1).T,atleast_2d(y1).T]),
        hstack([atleast_2d(x2).T,atleast_2d(y2).T]),
        ])
targets=concatenate([zeros(N),ones(N)])
target_names=['center','around']
feature_names=['x','y']



In [33]:

    
data=Struct(vectors=vectors,targets=targets,
                target_names=target_names,feature_names=feature_names)



In [34]:

    
C=Perceptron()
C.fit(data.vectors,data.targets)
print(("On Training Set:",C.percent_correct(data.vectors,data.targets)))
plot2D(data,classifier=C,axis_range=[-2,2,-2,2])









    



('On Training Set:', 56.666666666666664)



In [42]:

    
C=BackProp(hidden_layer_sizes = [6],max_iter=10000,tol=1e-4)
C.fit(data.vectors,data.targets)
print(("On Training Set:",C.percent_correct(data.vectors,data.targets)))
plot2D(data,classifier=C,axis_range=[-2,2,-2,2])









    



('On Training Set:', 98.333333333333329)



In [43]:

    
C=kNearestNeighbor()
C.fit(data.vectors,data.targets)
print(("On Training Set:",C.percent_correct(data.vectors,data.targets)))
plot2D(data,classifier=C,axis_range=[-2,2,-2,2])









    



('On Training Set:', 96.666666666666671)



In [44]:

    
C=CSC()
C.fit(data.vectors,data.targets)
print(("On Training Set:",C.percent_correct(data.vectors,data.targets)))
C.plot_centers()
plot2D(data,classifier=C,axis_range=[-2,2,-2,2])









    



('On Training Set:', 100.0)

8x8 - Autoencoder



In [45]:

    
vectors=eye(8)
targets=arange(1,9)
print((vectors,targets))









    



(array([[ 1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.]]), array([1, 2, 3, 4, 5, 6, 7, 8]))



In [47]:

    
C=BackProp(activation='logistic',hidden_layer_sizes = [3],max_iter=10000,tol=1e-4)
C.fit(vectors,targets)
print((C.predict(vectors)))









    



[1 2 3 4 5 6 7 8]



In [48]:

    
h,y=C.output(vectors)



In [49]:

    
h









    Out[49]:





array([[ 0.95188517,  0.95754607,  0.0190463 ],
       [ 0.95842582,  0.01367441,  0.93215547],
       [ 0.95640847,  0.9240123 ,  0.9521281 ],
       [ 0.04617869,  0.89309455,  0.02549217],
       [ 0.02920632,  0.01474587,  0.11884015],
       [ 0.95592761,  0.16780924,  0.01879628],
       [ 0.02321577,  0.01915328,  0.96401543],
       [ 0.02920981,  0.93431338,  0.93549868]])



In [50]:

    
h.round()









    Out[50]:





array([[ 1.,  1.,  0.],
       [ 1.,  0.,  1.],
       [ 1.,  1.,  1.],
       [ 0.,  1.,  0.],
       [ 0.,  0.,  0.],
       [ 1.,  0.,  0.],
       [ 0.,  0.,  1.],
       [ 0.,  1.,  1.]])



In [51]:

    
y.round()









    Out[51]:





array([[ 1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.]])



In [52]:

    
C.predict(vectors)









    Out[52]:





array([1, 2, 3, 4, 5, 6, 7, 8])



In [53]:

    
y.shape









    Out[53]:





(8, 8)



In [54]:

    
imshow(h,interpolation='nearest',cmap=cm.gray)
colorbar()









    Out[54]:





<matplotlib.colorbar.Colorbar at 0x1158789e8>



In [56]:

    
weights_xh,weights_hy=C.weights



In [57]:

    
plot(weights_xh,'-o')









    Out[57]:





[<matplotlib.lines.Line2D at 0x1162aaf60>,
 <matplotlib.lines.Line2D at 0x1162aa9b0>,
 <matplotlib.lines.Line2D at 0x1162aacf8>]



In [58]:

    
plot(weights_hy,'-o')









    Out[58]:





[<matplotlib.lines.Line2D at 0x1198a8780>,
 <matplotlib.lines.Line2D at 0x1198a87f0>,
 <matplotlib.lines.Line2D at 0x1198a8f28>,
 <matplotlib.lines.Line2D at 0x1198a8e48>,
 <matplotlib.lines.Line2D at 0x1198a8a90>,
 <matplotlib.lines.Line2D at 0x1198a83c8>,
 <matplotlib.lines.Line2D at 0x1198a8be0>,
 <matplotlib.lines.Line2D at 0x1199069b0>]

Tuning the number of hidden units



In [59]:

    
data=load_excel('data/iris.xls')
data_train,data_test=split(data,test_size=0.75)









    



iris.data 151 5
150 vectors of length 4
Feature names: 'petal length in cm', 'petal width in cm', 'sepal length in cm', 'sepal width in cm'
Target values given.
Target names: 'Iris-setosa', 'Iris-versicolor', 'Iris-virginica'
Mean:  [ 3.75866667  1.19866667  5.84333333  3.054     ]
Median:  [ 4.35  1.3   5.8   3.  ]
Stddev:  [ 1.75852918  0.76061262  0.82530129  0.43214658]
Original vector shape:  (150, 4)
Train vector shape:  (37, 4)
Test vector shape:  (113, 4)

select which number of hidden units to use



In [61]:

    
hidden=list(range(1,10))
percent_correct=[]
for n in hidden:
    C=BackProp(hidden_layer_sizes = [n],tol=1e-4,max_iter=10000)
    C.fit(data_train.vectors,data_train.targets)
    percent_correct.append(C.percent_correct(data_test.vectors,data_test.targets))
    
plot(hidden,percent_correct,'-o')
xlabel('Number of Hidden Units')
ylabel('Percent Correct on Test Data')









    Out[61]:





<matplotlib.text.Text at 0x1153efac8>



In [ ]:



In [ ]:



In [ ]: