forked from rgrosse/compositional_structure_search
/
initialization.py
181 lines (132 loc) · 7.17 KB
/
initialization.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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
import numpy as np
nax = np.newaxis
from algorithms import low_rank_poisson, crp, ibp, sparse_coding, chains
import grammar
import observations
import recursive
from utils import misc
debugger = None
def init_low_rank(data_matrix, num_iter=200):
m, n = data_matrix.m, data_matrix.n
state, X = low_rank_poisson.fit_model(data_matrix, 2, num_iter=num_iter)
U, V, ssq_U, ssq_N = state.U, state.V, state.ssq_U, state.ssq_N
U /= ssq_U[nax, :] ** 0.25
V *= ssq_U[:, nax] ** 0.25
left = recursive.GaussianNode(U, 'col', np.sqrt(ssq_U))
right = recursive.GaussianNode(V, 'row', np.sqrt(ssq_U))
pred = np.dot(U, V)
X = data_matrix.sample_latent_values(pred, ssq_N)
noise = recursive.GaussianNode(X - pred, 'scalar', ssq_N)
return recursive.SumNode([recursive.ProductNode([left, right]), noise])
def init_row_clustering(data_matrix, isotropic, num_iter=200):
m, n = data_matrix.m, data_matrix.n
state = crp.fit_model(data_matrix, isotropic_w=isotropic, isotropic_b=isotropic, num_iter=num_iter)
U = np.zeros((m, state.assignments.max() + 1), dtype=int)
U[np.arange(m), state.assignments] = 1
left = recursive.MultinomialNode(U)
if isotropic:
right = recursive.GaussianNode(state.centers, 'scalar', state.sigma_sq_b)
else:
right = recursive.GaussianNode(state.centers, 'col', state.sigma_sq_b)
pred = state.centers[state.assignments, :]
X = data_matrix.sample_latent_values(pred, state.sigma_sq_w * np.ones((m, n)))
if isotropic:
noise = recursive.GaussianNode(X - pred, 'scalar', state.sigma_sq_w)
else:
noise = recursive.GaussianNode(X - pred, 'col', state.sigma_sq_w)
return recursive.SumNode([recursive.ProductNode([left, right]), noise])
def init_col_clustering(data_matrix, isotropic, num_iter=200):
return init_row_clustering(data_matrix.transpose(), isotropic, num_iter=num_iter).transpose()
def init_row_binary(data_matrix, num_iter=200):
state = ibp.fit_model(data_matrix, num_iter=num_iter)
left = recursive.BernoulliNode(state.Z)
right = recursive.GaussianNode(state.A, 'scalar', state.sigma_sq_f)
pred = np.dot(state.Z, state.A)
X = data_matrix.sample_latent_values(pred, state.sigma_sq_n)
noise = recursive.GaussianNode(X - pred, 'scalar', state.sigma_sq_n)
return recursive.SumNode([recursive.ProductNode([left, right]), noise])
def init_col_binary(data_matrix, num_iter=200):
return init_row_binary(data_matrix.transpose(), num_iter=num_iter).transpose()
def init_row_chain(data_matrix, num_iter=200):
states, sigma_sq_D, sigma_sq_N = chains.fit_model(data_matrix, num_iter=num_iter)
integ = chains.integration_matrix(data_matrix.m_orig)[data_matrix.row_ids, :]
left = recursive.IntegrationNode(integ)
temp = np.vstack([states[0, :][nax, :],
states[1:, :] - states[:-1, :]])
right = recursive.GaussianNode(temp, 'scalar', sigma_sq_D)
pred = states[data_matrix.row_ids, :]
X = data_matrix.sample_latent_values(pred, sigma_sq_N)
noise = recursive.GaussianNode(X - pred, 'scalar', sigma_sq_N)
return recursive.SumNode([recursive.ProductNode([left, right]), noise])
def init_col_chain(data_matrix, num_iter=200):
return init_row_chain(data_matrix.transpose(), num_iter=num_iter).transpose()
def init_sparsity(data_matrix, mu_Z_mode, num_iter=200):
if mu_Z_mode == 'row':
return init_sparsity(data_matrix.transpose(), 'col', num_iter).transpose()
elif mu_Z_mode == 'col':
by_column = True
elif mu_Z_mode == 'scalar':
by_column = False
# currently, data_matrix should always be real-valued with no missing values, so this just
# passes on data_matrix.observations.values; we may want to replace it with interval observations
# obtained from slice sampling
S = data_matrix.sample_latent_values(np.zeros((data_matrix.m, data_matrix.n)),
np.ones((data_matrix.m, data_matrix.n)))
Z = np.random.normal(-1., 1., size=S.shape)
# sparse_coding.py wants a full sparse coding problem, so pass in None for the things
# that aren't relevant here
state = sparse_coding.SparseCodingState(S, None, Z, None, -1., 1., None)
pbar = misc.pbar(num_iter)
for i in range(num_iter):
sparse_coding.sample_Z(state)
state.mu_Z = sparse_coding.cond_mu_Z(state, by_column).sample()
state.sigma_sq_Z = sparse_coding.cond_sigma_sq_Z(state).sample()
if hasattr(debugger, 'after_init_sparsity_iter'):
debugger.after_init_sparsity_iter(locals())
pbar.update(i)
pbar.finish()
scale_node = recursive.GaussianNode(state.Z, 'scalar', state.sigma_sq_Z)
return recursive.GSMNode(state.S, scale_node, mu_Z_mode, state.mu_Z)
def initialize(data_matrix, root, old_structure, new_structure, num_iter=200):
root = root.copy()
if old_structure == new_structure:
return root
node, old_dist, rule = recursive.find_changed_node(root, old_structure, new_structure)
old = root.value()
# if we're replacing the root, pass on the observation model; otherwise, treat
# the node we're factorizing as exact real-valued observations
if node is root:
inner_data_matrix = data_matrix
else:
row_ids = recursive.row_ids_for(data_matrix, node)
col_ids = recursive.col_ids_for(data_matrix, node)
m_orig, n_orig = recursive.orig_shape_for(data_matrix, node)
frv = observations.DataMatrix.from_real_values
inner_data_matrix = frv(node.value(), row_ids=row_ids, col_ids=col_ids,
m_orig=m_orig, n_orig=n_orig)
print 'Initializing %s from %s...' % (grammar.pretty_print(new_structure), grammar.pretty_print(old_structure))
if rule == grammar.parse("gg+g"):
new_node = init_low_rank(inner_data_matrix, num_iter=num_iter)
elif rule == grammar.parse("mg+g"):
isotropic = (node is root)
new_node = init_row_clustering(inner_data_matrix, isotropic, num_iter=num_iter)
elif rule == grammar.parse("gM+g"):
isotropic = (node is root)
new_node = init_col_clustering(inner_data_matrix, isotropic, num_iter=num_iter)
elif rule == grammar.parse("bg+g"):
new_node = init_row_binary(inner_data_matrix, num_iter=num_iter)
elif rule == grammar.parse("gB+g"):
new_node = init_col_binary(inner_data_matrix, num_iter=num_iter)
elif rule == grammar.parse("cg+g"):
new_node = init_row_chain(inner_data_matrix, num_iter=num_iter)
elif rule == grammar.parse("gC+g"):
new_node = init_col_chain(inner_data_matrix, num_iter=num_iter)
elif rule == grammar.parse("s(g)"):
new_node = init_sparsity(inner_data_matrix, node.variance_type, num_iter=num_iter)
else:
raise RuntimeError('Unknown production rule: %s ==> %s' % (grammar.pretty_print(old_dist),
grammar.pretty_print(rule)))
root = recursive.splice(root, node, new_node)
if isinstance(data_matrix.observations, observations.RealObservations):
assert np.allclose(root.value()[data_matrix.observations.mask], old[data_matrix.observations.mask])
return root