forked from blt2114/overpruning_in_variational_bnns
-
Notifications
You must be signed in to change notification settings - Fork 0
/
synthetic_data.py
94 lines (83 loc) · 3.55 KB
/
synthetic_data.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
import tensorflow as tf
from scipy import stats
import numpy as np
def gen_synthetic_data(dim=2, n_pts=10000, bimodal=False, heteroscedastic=True,
asymetric=False):
"""gen_synthetic_data generates synthetic data with 1D output which is
bimodal and heteroscedastic.
Args:
dim: dimensionality of the input
n_pts: number of points to generate
bimodal: true to generate bimodal data
heteroscedastic: true to generate heteroscedastic data
asymetric: set True to have noise have asymetric tails
"""
### Generate fake data
bounds = [-np.pi,np.pi]
range_size = bounds[1]-bounds[0]
if heteroscedastic:
noise_scale = 3. # for noise
else:
noise_scale = 0. # for noise
global_noise = 1.0 # additional homoscedastic gaussian noise
signal_scale = 5.
if bimodal:
n_pts_mode = int(n_pts/2.)
else:
n_pts_mode = n_pts
# Generate X for first half of data
X = np.random.uniform(bounds[0],bounds[1],size=[n_pts_mode,dim])
Y_std = noise_scale*abs(np.sin(X).prod(axis=1))+global_noise # Heteroscedastic noise
if asymetric:
Y = abs(np.random.normal(0.,abs(Y_std)))+signal_scale*np.sin(X).prod(axis=1)
else:
Y = np.random.normal(0.,abs(Y_std))+signal_scale*np.sin(X).prod(axis=1)
# Generate data from second mode
if bimodal:
X_more = np.random.uniform(bounds[0],bounds[1],size=[n_pts_mode,dim])
Y_std_more = noise_scale*abs(np.sin(X_more)).prod(axis=1)+global_noise
# The bimodality arises from using 'abs(X_more)' rather than simply
# X_more within sin
if asymetric:
Y_more = abs(np.random.normal(0., abs(Y_std_more)))+signal_scale*np.sin(abs(X_more).prod(axis=1))
else:
Y_more = np.random.normal(0., abs(Y_std_more))+signal_scale*np.sin(abs(X_more).prod(axis=1))
# concatenate two datasets together for bimodal signal
X, Y = np.array(list(X)+list(X_more)), np.array(list(Y)+list(Y_more))
Y = Y.reshape([n_pts, 1])
return X, Y
#### We define the parameters of the mixture globally for now.
n_clust = 4
prior_mu_std = 2. # was 1.
prior_std_scale = 5.5
n_samples = 10000
mix_props = np.random.dirichlet([18.5]*n_clust)# mixing proportions
mus = np.random.normal(loc=[0.,0.],scale=prior_mu_std,size=[n_clust,2])
cov_mats = np.random.gamma(prior_std_scale, scale=0.05,size=[n_clust,2,2])
cov_mats[:,0,0] = np.random.uniform(low=0.7,high=1.4,size=[n_clust])*cov_mats[:,1,1]
cov_mats[:,1,0] = np.random.uniform(low=0.2,high=0.4,size=[n_clust])*cov_mats[:,1,1]
def gen_mog_synthetic_data():
"""gen_mog_synthetic_data generates data from a 2D mixture of Gaussians.
"""
cov_mats[:,0,1]=cov_mats[:,1,0]
#print("mus: ",mus)
#print("cov_mats: ",cov_mats)
#print("mix_props: ",mix_props)
Y = []
for (mix_props_i, mu_i,cov_mat_i) in zip(mix_props,mus,cov_mats):
Y.extend(np.random.multivariate_normal(mu_i,cov_mat_i,size=[int(n_samples*mix_props_i)]))
Y = np.array(Y)
print("Y_shape",Y.shape)
Y = np.array(Y).reshape([len(Y[:,0]),2])
## at this point we don't have any inputs, so X is an empty array.
X = np.zeros([Y.shape[0],1], dtype=np.float32)
return X, Y
def p_mixture(y):
"""p_mixture calculates the probility of samples provied undert the
true, data generating mixture distribution.
Returns:
A vector of probabilities of the samples.
"""
return sum(stats.multivariate_normal.pdf(
x=y,mean=mu_i, cov= cov_i
)*mix_props_i for (mix_props_i, mu_i,cov_i) in zip(mix_props,mus,cov_mats))