forked from HIPS/autograd
-
Notifications
You must be signed in to change notification settings - Fork 0
/
mixture_variational_inference.py
167 lines (134 loc) · 6.74 KB
/
mixture_variational_inference.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
# Implements black-box variational inference, where the variational
# distribution is a mixture of Gaussians.
#
# This trick was written up by Alex Graves in this note:
# http://arxiv.org/abs/1607.05690
from __future__ import absolute_import
from __future__ import print_function
import matplotlib.pyplot as plt
import autograd.numpy as np
import autograd.numpy.random as npr
import autograd.scipy.stats.norm as norm
from autograd.scipy.special import logsumexp
from autograd import grad
from autograd.misc.optimizers import adam
def diag_gaussian_log_density(x, mu, log_std):
return np.sum(norm.logpdf(x, mu, np.exp(log_std)), axis=-1)
def unpack_gaussian_params(params):
# Variational dist is a diagonal Gaussian.
D = np.shape(params)[0] // 2
mean, log_std = params[:D], params[D:]
return mean, log_std
def variational_log_density_gaussian(params, x):
mean, log_std = unpack_gaussian_params(params)
return diag_gaussian_log_density(x, mean, log_std)
def sample_diag_gaussian(params, num_samples, rs):
mean, log_std = unpack_gaussian_params(params)
D = np.shape(mean)[0]
return rs.randn(num_samples, D) * np.exp(log_std) + mean
def variational_lower_bound(params, t, logprob, sampler, log_density,
num_samples, rs):
"""Provides a stochastic estimate of the variational lower bound,
for any variational family and model density."""
samples = sampler(params, num_samples, rs)
log_qs = log_density(params, samples)
log_ps = logprob(samples, t)
log_ps = np.reshape(log_ps, (num_samples, -1))
log_qs = np.reshape(log_qs, (num_samples, -1))
return np.mean(log_ps - log_qs)
def init_gaussian_var_params(D, mean_mean=-1, log_std_mean=-5,
scale=0.1, rs=npr.RandomState(0)):
init_mean = mean_mean * np.ones(D) + rs.randn(D) * scale
init_log_std = log_std_mean * np.ones(D) + rs.randn(D) * scale
return np.concatenate([init_mean, init_log_std])
def log_normalize(x):
return x - logsumexp(x)
def build_mog_bbsvi(logprob, num_samples, k=10, rs=npr.RandomState(0)):
init_component_var_params = init_gaussian_var_params
component_log_density = variational_log_density_gaussian
component_sample = sample_diag_gaussian
def unpack_mixture_params(mixture_params):
log_weights = log_normalize(mixture_params[:k])
var_params = np.reshape(mixture_params[k:], (k, -1))
return log_weights, var_params
def init_var_params(D, rs=npr.RandomState(0), **kwargs):
log_weights = np.ones(k)
component_weights = [init_component_var_params(D, rs=rs, **kwargs) for i in range(k)]
return np.concatenate([log_weights] + component_weights)
def sample(var_mixture_params, num_samples, rs):
"""Sample locations aren't a continuous function of parameters
due to multinomial sampling."""
log_weights, var_params = unpack_mixture_params(var_mixture_params)
samples = np.concatenate([component_sample(params_k, num_samples, rs)[:, np.newaxis, :]
for params_k in var_params], axis=1)
ixs = np.random.choice(k, size=num_samples, p=np.exp(log_weights))
return np.array([samples[i, ix, :] for i, ix in enumerate(ixs)])
def mixture_log_density(var_mixture_params, x):
"""Returns a weighted average over component densities."""
log_weights, var_params = unpack_mixture_params(var_mixture_params)
component_log_densities = np.vstack([component_log_density(params_k, x)
for params_k in var_params]).T
return logsumexp(component_log_densities + log_weights, axis=1, keepdims=False)
def mixture_elbo(var_mixture_params, t):
# We need to only sample the continuous component parameters,
# and integrate over the discrete component choice
def mixture_lower_bound(params):
"""Provides a stochastic estimate of the variational lower bound."""
samples = component_sample(params, num_samples, rs)
log_qs = mixture_log_density(var_mixture_params, samples)
log_ps = logprob(samples, t)
log_ps = np.reshape(log_ps, (num_samples, -1))
log_qs = np.reshape(log_qs, (num_samples, -1))
return np.mean(log_ps - log_qs)
log_weights, var_params = unpack_mixture_params(var_mixture_params)
component_elbos = np.stack(
[mixture_lower_bound(params_k) for params_k in var_params])
return np.sum(component_elbos*np.exp(log_weights))
return init_var_params, mixture_elbo, mixture_log_density, sample
if __name__ == '__main__':
# Specify an inference problem by its unnormalized log-density.
D = 2
def log_density(x, t):
mu, log_sigma = x[:, 0], x[:, 1]
sigma_density = norm.logpdf(log_sigma, 0, 1.35)
mu_density = norm.logpdf(mu, -0.5, np.exp(log_sigma))
sigma_density2 = norm.logpdf(log_sigma, 0.1, 1.35)
mu_density2 = norm.logpdf(mu, 0.5, np.exp(log_sigma))
return np.logaddexp(sigma_density + mu_density,
sigma_density2 + mu_density2)
init_var_params, elbo, variational_log_density, variational_sampler = \
build_mog_bbsvi(log_density, num_samples=40, k=10)
def objective(params, t):
return -elbo(params, t)
# Set up plotting code
def plot_isocontours(ax, func, xlimits=[-2, 2], ylimits=[-4, 2],
numticks=101, cmap=None):
x = np.linspace(*xlimits, num=numticks)
y = np.linspace(*ylimits, num=numticks)
X, Y = np.meshgrid(x, y)
zs = func(np.concatenate([np.atleast_2d(X.ravel()), np.atleast_2d(Y.ravel())]).T)
Z = zs.reshape(X.shape)
plt.contour(X, Y, Z, cmap=cmap)
ax.set_yticks([])
ax.set_xticks([])
fig = plt.figure(figsize=(8,8), facecolor='white')
ax = fig.add_subplot(111, frameon=False)
plt.ion()
plt.show(block=False)
num_plotting_samples = 51
def callback(params, t, g):
print("Iteration {} lower bound {}".format(t, -objective(params, t)))
plt.cla()
target_distribution = lambda x: np.exp(log_density(x, t))
var_distribution = lambda x: np.exp(variational_log_density(params, x))
plot_isocontours(ax, target_distribution)
plot_isocontours(ax, var_distribution, cmap=plt.cm.bone)
ax.set_autoscale_on(False)
rs = npr.RandomState(0)
samples = variational_sampler(params, num_plotting_samples, rs)
plt.plot(samples[:, 0], samples[:, 1], 'x')
plt.draw()
plt.pause(1.0/30.0)
print("Optimizing variational parameters...")
variational_params = adam(grad(objective), init_var_params(D), step_size=0.1,
num_iters=2000, callback=callback)