/
diff_lvq_softmax.py
151 lines (114 loc) · 4.85 KB
/
diff_lvq_softmax.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
import theano
import theano.tensor as tensor
import numpy
num_protos = 3
n_out= 2
n_examples = 500
eta=0.2
h3 = tensor.fmatrix('x')
mask = tensor.fmatrix('mask')
W = theano.shared(numpy.random.normal(size=(num_protos, n_out)))
D = ((W**2).sum(axis=1, keepdims=True).T + (h3**2).sum(axis=1, keepdims=True) - 2*tensor.dot(h3, W.T))
d_correct = (D + (1-mask)*numpy.float32(1e30)).min(axis=1)
d_incorrect = (D + mask*numpy.float32(1e30)).min(axis=1)
mu = numpy.float32(0.2)
cost = (d_correct - d_incorrect)/(d_correct + d_incorrect)
misclass = (tensor.switch(d_correct - d_incorrect < 0, 0.0, 1.0).sum())/mask.shape[0]
loss = (tensor.exp(2.0*cost)).mean()
params = [W]
grad = theano.grad(loss, params)
updates = [p-eta*g_p for p, g_p in zip(params, grad)]
f = theano.function([h3, mask], [loss, misclass], updates=zip(params, updates))
l = theano.function([h3, mask], [loss, misclass])
f_D = theano.function([h3], D)
d1 = numpy.random.multivariate_normal(mean=(0, -6), cov=numpy.eye(2), size=n_examples).astype('float32')
m1 = numpy.zeros((n_examples, num_protos), 'float32')
m1[:, 0] = 1
d2 = numpy.random.multivariate_normal(mean=(-3, 0), cov=numpy.eye(2), size=n_examples).astype('float32')
m2 = numpy.zeros((n_examples, num_protos), 'float32')
m2[:, 1] = 1
d3 = numpy.random.multivariate_normal(mean=(3, 0), cov=numpy.eye(2), size=n_examples).astype('float32')
m3 = numpy.zeros((n_examples, num_protos), 'float32')
m3[:, 2] = 1
all_labels = numpy.concatenate((numpy.zeros((n_examples,), 'int32'), numpy.ones((n_examples,), 'int32'), numpy.ones((n_examples,), 'int32')*2))
all_data = numpy.vstack((d1, d2, d3))
all_mask = numpy.vstack((m1, m2, m3))
ind = numpy.random.permutation(num_protos*n_examples)
data = all_data[ind, :]
mask = all_mask[ind, :]
labels = all_labels[ind]
batch_size= 200
n_batches = (num_protos*n_examples)/batch_size
for j in range(100):
for i in range(n_batches):
if i == 0:
print f(data[i*batch_size:(i+1)*batch_size, :], mask[i*batch_size:(i+1)*batch_size, :])
else:
f(data[i*batch_size:(i+1)*batch_size, :], mask[i*batch_size:(i+1)*batch_size, :])
import matplotlib.pyplot as plt
from matplotlib.axes import Axes
import matplotlib.cm as cm
from scipy.spatial import voronoi_plot_2d, Voronoi
weights = W.get_value()
#vor = Voronoi(weights)
#voronoi_plot_2d(vor)
plt.plot(d1[:, 0], d1[:, 1], 'ro', alpha=0.5, markersize=3)
plt.plot(d2[:, 0], d2[:, 1], 'bo', alpha=0.5, markersize=3)
plt.plot(d3[:, 0], d3[:, 1], 'go', alpha=0.5, markersize=3)
plt.plot(weights[0, 0], weights[0, 1], 'r*', alpha=0.5, markersize=20)
plt.plot(weights[1, 0], weights[1, 1], 'b*', alpha=0.5, markersize=20)
plt.plot(weights[2, 0], weights[2, 1], 'g*', alpha=0.5, markersize=20)
X = numpy.arange(-10, 10, 0.1)
Y = numpy.arange(-10, 10, 0.1)
Z = numpy.zeros((len(X),len(Y)))
for i in range(len(X)):
for j in range(len(Y)):
inp = numpy.array([[X[i], Y[j]]]).astype('float32')
d = numpy.sort(f_D(inp))
Z[j, i] = -1*(d[0][0]- d[0][1])/(d[0][0] + d[0][1])
im = plt.imshow(Z[:, :], interpolation='bilinear', origin='lower',
cmap=cm.gray, extent=(-10, 10, -10, 10))
plt.colorbar(im, orientation='horizontal', shrink=0.8)
plt.xlim((-10, 10))
plt.ylim((-10, 10))
plt.savefig('mnist/figures/lvq.pdf')
###################
#### Softmax
###################
from blocks.bricks import Softmax
from blocks.bricks.cost import MisclassificationRate
W2 = theano.shared(numpy.random.normal(size=(n_out, num_protos)).astype('float32'))
b = theano.shared(numpy.zeros((num_protos,)).astype('float32'))
y = tensor.ivector('y')
h = tensor.dot(h3, W2) + b
h = tensor.switch(h < 0, -h , h)
sm = Softmax()
pred = sm.apply(h)
misclass = MisclassificationRate().apply(y, pred)
c = sm.categorical_cross_entropy(y, h).mean()
s_params = [W2, b]
s_grad = theano.grad(c, s_params)
s_updates = [p - numpy.float32(0.05)*g for p, g in zip(s_params, s_grad)]
s_f = theano.function([h3, y], [c, misclass], updates=zip(s_params, s_updates))
s_pred = theano.function([h3], pred)
for j in range(200):
for i in range(n_batches):
if i == 0:
print s_f(data[i*batch_size:(i+1)*batch_size, :], labels[i*batch_size:(i+1)*batch_size])
else:
s_f(data[i*batch_size:(i+1)*batch_size, :], labels[i*batch_size:(i+1)*batch_size])
print b.get_value()
fig1 = plt.figure()
plt.plot(d1[:, 0], d1[:, 1], 'ro', alpha=0.5, markersize=3)
plt.plot(d2[:, 0], d2[:, 1], 'bo', alpha=0.5, markersize=3)
plt.plot(d3[:, 0], d3[:, 1], 'go', alpha=0.5, markersize=3)
for i in range(len(X)):
for j in range(len(Y)):
inp = numpy.array([[X[i], Y[j]]]).astype('float32')
Z[j, i] = s_pred(inp).max()
im = plt.imshow(Z[:, :], interpolation='bilinear', origin='lower',
cmap=cm.gray, extent=(-10, 10, -10, 10))
plt.colorbar(im, orientation='horizontal', shrink=0.8)
plt.xlim((-10, 10))
plt.ylim((-10, 10))
plt.savefig('mnist/figures/softmax.pdf')