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
|
from numpy.random import normal
import numpy as np
from pickle import load
from random import shuffle
def init_vectors(K, sigma=0.01):
U = normal(0, sigma, size=(n+1, K))
V = normal(0, sigma, size=(m+1, K))
return U, V
def init_vectors_bis(K, sigma=0.01):
U = normal(0, sigma, size=(n+1, K))
V = normal(0, sigma, size=(m+1, K))
a = np.ones(n+1)
b = np.ones(m+1)
g = 1
return U, V, a, b, g
def sgd_step(U, V, lamb=0.05, sigmas=1.0):
keys = train.keys()
shuffle(keys)
for (i, j) in keys:
r = train[(i, j)]
ui = np.copy(U[i])
vj = np.copy(V[j])
a = (lamb / sigmas) * (float(r) - np.inner(ui, vj))
U[i] += a * vj
V[j] += a * ui
def sgd_step_bis(U, V, a, b, g, lamb=0.05, sigmas=1.0):
keys = train.keys()
shuffle(keys)
for (i, j) in keys:
r = train[(i, j)]
ui = np.copy(U[i])
vj = np.copy(V[j])
e = (lamb / sigmas) * (float(r) - np.inner(ui, vj) - a[i] - b[j] - g)
U[i] += e * vj
V[j] += e * ui
a[i] += e
b[j] += e
g += e
return g
def mse(U, V, d):
serror = sum((float(r) - np.inner(U[i], V[j])) ** 2
for (i, j), r in d.iteritems())
return serror / len(d)
def mse_bis(U, V, a, b, g, d):
serror = sum((float(r) - np.inner(U[i], V[j]) - a[i] - b[j] - g) ** 2
for (i, j), r in d.iteritems())
return serror / len(d)
def part_c():
for K in xrange(1, 11):
U, V = init_vectors(K)
for t in xrange(10):
sgd_step(U, V)
print K, mse(U, V, train), mse(U, V, test)
def part_d():
U, V, a, b, g = init_vectors_bis(2)
for t in xrange(10):
print t
g = sgd_step_bis(U, V, a, b, g)
print g
i = range(len(b))
mi = min(i, key=lambda x: b[x])
ma = max(i, key=lambda x: b[x])
print mi, b[mi]
print ma, b[ma]
print mse_bis(U, V, a, b, g, train), mse_bis(U, V, a, b, g, test)
if __name__ == "__main__":
n, m, train, test = load(open("data.pickle", "rb"))
part_d()
|
