/
ebcc.py
204 lines (178 loc) · 7.4 KB
/
ebcc.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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
from typing import Optional, Any, Union
import numpy as np
import scipy.sparse as ssp
from scipy.special import digamma, gammaln
from scipy.stats import entropy, dirichlet
from tqdm import trange
from wrench.basemodel import BaseLabelModel
from wrench.dataset import BaseDataset
from ..utils import create_tuples
def ebcc_vb(tuples,
num_items,
num_workers,
num_classes,
a_pi=0.1,
num_groups=10, # M
alpha=1, # alpha_k, it can be 1 or \sum_i gamma_ik
a_v=4, # beta_kk
b_v=1, # beta_kk', k neq k'
eta_km=None,
nu_k=None,
mu_jkml=None,
eval=False,
seed=1234,
inference_iter=500,
empirical_prior=False):
y_is_one_lij = []
y_is_one_lji = []
for k in range(num_classes):
selected = (tuples[:, 2] == k)
coo_ij = ssp.coo_matrix((np.ones(selected.sum()),
tuples[selected, :2].T),
shape=(num_items, num_workers),
dtype=np.bool)
y_is_one_lij.append(coo_ij.tocsr())
y_is_one_lji.append(coo_ij.T.tocsr())
beta_kl = np.eye(num_classes) * (a_v - b_v) + b_v
# initialize z_ik, zg_ikm, c_ik, gamma_ik, sigma_ik
z_ik = np.zeros((num_items, num_classes))
for l in range(num_classes):
z_ik[:, [l]] += y_is_one_lij[l].sum(axis=-1) + 1e-8
z_ik /= z_ik.sum(axis=-1, keepdims=True)
if empirical_prior:
alpha = z_ik.sum(axis=0)
np.random.seed(seed)
zg_ikm = np.random.dirichlet(np.ones(num_groups), z_ik.shape) * z_ik[:, :, None]
for it in range(inference_iter):
if eval is False:
eta_km = a_pi / num_groups + zg_ikm.sum(axis=0)
nu_k = alpha + z_ik.sum(axis=0)
mu_jkml = np.zeros((num_workers, num_classes, num_groups, num_classes)) + beta_kl[None, :, None, :]
for l in range(num_classes):
for k in range(num_classes):
mu_jkml[:, k, :, l] += y_is_one_lji[l].dot(zg_ikm[:, k, :])
Eq_log_pi_km = digamma(eta_km) - digamma(eta_km.sum(axis=-1, keepdims=True))
Eq_log_tau_k = digamma(nu_k) - digamma(nu_k.sum())
Eq_log_v_jkml = digamma(mu_jkml) - digamma(mu_jkml.sum(axis=-1, keepdims=True))
zg_ikm[:] = Eq_log_pi_km[None, :, :] + Eq_log_tau_k[None, :, None]
for l in range(num_classes):
for k in range(num_classes):
zg_ikm[:, k, :] += y_is_one_lij[l].dot(Eq_log_v_jkml[:, k, :, l])
zg_ikm = np.exp(zg_ikm)
zg_ikm /= zg_ikm.reshape(num_items, -1).sum(axis=-1)[:, None, None]
last_z_ik = z_ik
z_ik = zg_ikm.sum(axis=-1)
if np.allclose(last_z_ik, z_ik, atol=1e-3):
break
ELBO = ((eta_km - 1) * Eq_log_pi_km).sum() + ((nu_k - 1) * Eq_log_tau_k).sum() + (
(mu_jkml - 1) * Eq_log_v_jkml).sum()
ELBO += dirichlet.entropy(nu_k)
for k in range(num_classes):
ELBO += dirichlet.entropy(eta_km[k])
ELBO += (gammaln(mu_jkml) - (mu_jkml - 1) * digamma(mu_jkml)).sum()
alpha0_jkm = mu_jkml.sum(axis=-1)
ELBO += ((alpha0_jkm - num_classes) * digamma(alpha0_jkm) - gammaln(alpha0_jkm)).sum()
ELBO += entropy(zg_ikm.reshape(num_items, -1).T).sum()
return z_ik, ELBO, eta_km, nu_k, mu_jkml
class EBCC(BaseLabelModel):
"""Enhanced BCC (eBCC)
Usage:
ebcc = EBCC(num_groups, a_pi, a_v, b_v, repeat, inference_iter, empirical_prior)
ebcc.fit(train_data)
ebcc.test(test_data)
Parameters:
num_groups: number of subtypes
a_pi: The parameter of dirichlet distribution to generate mixture weight.
a_v: b_kk, number of corrected labeled items under every class.
b_v: b_kk', all kind of miss has made b_kk' times.
repeat: ELBO update times.
inference_iter: Iterations of variational inference.
empirical_prior: The empirical prior of alpha.
seed: Random seed.
"""
def __init__(self,
num_groups: Optional[int] = 10,
a_pi: Optional[float] = 0.1,
alpha: Optional[float] = 1,
a_v: Optional[float] = 4,
b_v: Optional[float] = 1,
repeat: Optional[int] = 5,
inference_iter: Optional[int] = 1000,
seed: Optional[int] = None,
empirical_prior=True,
**kwargs: Any):
super().__init__()
self.hyperparas = {
'num_groups': num_groups,
'a_pi': a_pi,
'alpha': alpha,
'a_v': a_v,
'b_v': b_v,
'empirical_prior': empirical_prior,
'inference_iter': inference_iter,
**kwargs
}
self.params = {
'seed': None,
'eta_km': None,
'nu_k': None,
'mu_jkml': None,
}
self.seed = seed
self.repeat = repeat
def fit(self,
dataset_train: Union[BaseDataset, np.ndarray],
dataset_valid: Optional[Union[BaseDataset, np.ndarray]] = None,
y_valid: Optional[np.ndarray] = None,
n_class: Optional[int] = None,
verbose: Optional[bool] = False,
*args: Any,
**kwargs: Any):
tuples = create_tuples(dataset_train)
num_items, _, num_classes = tuples.max(axis=0) + 1
num_workers = len(dataset_train.weak_labels[0])
max_elbo = float('-inf')
if self.seed is None:
for _ in trange(0, self.repeat, unit='epoch'):
seed = np.random.randint(1e8)
prediction, elbo, p1, p2, p3 = ebcc_vb(tuples,
num_items, num_workers, num_classes,
seed=seed,
**self.hyperparas)
if elbo > max_elbo:
print(f'update elbo: new: {elbo}, old: {max_elbo}')
self.params = {
'seed': seed,
'eta_km': p1,
'nu_k': p2,
'mu_jkml': p3
}
max_elbo = elbo
pred = prediction
else:
pred, elbo, p1, p2, p3 = ebcc_vb(tuples,
num_items, num_workers, num_classes,
seed=self.seed,
**self.hyperparas)
self.params = {
'seed': self.seed,
'eta_km': p1,
'nu_k': p2,
'mu_jkml': p3
}
return pred
def predict_proba(self,
dataset: Union[BaseDataset, np.ndarray],
**kwargs: Any):
tuples = create_tuples(dataset)
num_items, _, num_classes = tuples.max(axis=0) + 1
num_workers = len(dataset.weak_labels[0])
eval = True
if self.params['nu_k'] is None or self.params['mu_jkml'] is None:
eval = False
pred, elbo, _, _, _ = ebcc_vb(tuples,
num_items, num_workers, num_classes,
eval=eval,
**self.hyperparas,
**self.params)
return pred