-
-
Notifications
You must be signed in to change notification settings - Fork 986
/
rejection_gamma.py
198 lines (167 loc) · 8.39 KB
/
rejection_gamma.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
import torch
from pyro.distributions.rejector import Rejector
from pyro.distributions.score_parts import ScoreParts
from pyro.distributions.torch import Beta, Dirichlet, Gamma, Normal
from pyro.distributions.util import copy_docs_from, weakmethod
@copy_docs_from(Gamma)
class RejectionStandardGamma(Rejector):
"""
Naive Marsaglia & Tsang rejection sampler for standard Gamma distibution.
This assumes `concentration >= 1` and does not boost `concentration` or augment shape.
"""
def __init__(self, concentration):
if concentration.data.min() < 1:
raise NotImplementedError('concentration < 1 is not supported')
self.concentration = concentration
self._standard_gamma = Gamma(concentration, concentration.new([1.]).squeeze().expand_as(concentration))
# The following are Marsaglia & Tsang's variable names.
self._d = self.concentration - 1.0 / 3.0
self._c = 1.0 / torch.sqrt(9.0 * self._d)
# Compute log scale using Gamma.log_prob().
x = self._d.detach() # just an arbitrary x.
log_scale = self.propose_log_prob(x) + self.log_prob_accept(x) - self.log_prob(x)
super(RejectionStandardGamma, self).__init__(self.propose, self.log_prob_accept, log_scale)
def expand(self, batch_shape, _instance=None):
new = self._get_checked_instance(RejectionStandardGamma, _instance)
batch_shape = torch.Size(batch_shape)
new.concentration = self.concentration.expand(batch_shape)
new._standard_gamma = self._standard_gamma.expand(batch_shape)
new._d = self._d.expand(batch_shape)
new._c = self._c.expand(batch_shape)
# Compute log scale using Gamma.log_prob().
x = new._d.detach() # just an arbitrary x.
log_scale = new.propose_log_prob(x) + new.log_prob_accept(x) - new.log_prob(x)
super(RejectionStandardGamma, new).__init__(new.propose, new.log_prob_accept, log_scale)
new._validate_args = self._validate_args
return new
@weakmethod
def propose(self, sample_shape=torch.Size()):
# Marsaglia & Tsang's x == Naesseth's epsilon`
x = torch.randn(sample_shape + self.concentration.shape,
dtype=self.concentration.dtype,
device=self.concentration.device)
y = 1.0 + self._c * x
v = y * y * y
return (self._d * v).clamp_(1e-30, 1e30)
def propose_log_prob(self, value):
v = value / self._d
result = -self._d.log()
y = v.pow(1 / 3)
result -= torch.log(3 * y ** 2)
x = (y - 1) / self._c
result -= self._c.log()
result += Normal(torch.zeros_like(self.concentration), torch.ones_like(self.concentration)).log_prob(x)
return result
@weakmethod
def log_prob_accept(self, value):
v = value / self._d
y = torch.pow(v, 1.0 / 3.0)
x = (y - 1.0) / self._c
log_prob_accept = 0.5 * x * x + self._d * (1.0 - v + torch.log(v))
log_prob_accept[y <= 0] = -float('inf')
return log_prob_accept
def log_prob(self, x):
return self._standard_gamma.log_prob(x)
@copy_docs_from(Gamma)
class RejectionGamma(Gamma):
has_rsample = True
def __init__(self, concentration, rate, validate_args=None):
super(RejectionGamma, self).__init__(concentration, rate, validate_args=validate_args)
self._standard_gamma = RejectionStandardGamma(concentration)
self.rate = rate
def expand(self, batch_shape, _instance=None):
new = self._get_checked_instance(RejectionGamma, _instance)
new = super(RejectionGamma, self).expand(batch_shape, new)
new._standard_gamma = self._standard_gamma.expand(batch_shape)
new._validate_args = self._validate_args
return new
def rsample(self, sample_shape=torch.Size()):
return self._standard_gamma.rsample(sample_shape) / self.rate
def log_prob(self, x):
return self._standard_gamma.log_prob(x * self.rate) + torch.log(self.rate)
def score_parts(self, x):
log_prob, score_function, _ = self._standard_gamma.score_parts(x * self.rate)
log_prob = log_prob + torch.log(self.rate)
return ScoreParts(log_prob, score_function, log_prob)
@copy_docs_from(Gamma)
class ShapeAugmentedGamma(Gamma):
"""
This implements the shape augmentation trick of
Naesseth, Ruiz, Linderman, Blei (2017) https://arxiv.org/abs/1610.05683
"""
has_rsample = True
def __init__(self, concentration, rate, boost=1, validate_args=None):
if concentration.min() + boost < 1:
raise ValueError('Need to boost at least once for concentration < 1')
super(ShapeAugmentedGamma, self).__init__(concentration, rate, validate_args=validate_args)
self.concentration = concentration
self._boost = boost
self._rejection_gamma = RejectionGamma(concentration + boost, rate)
self._unboost_x_cache = None, None
def expand(self, batch_shape, _instance=None):
new = self._get_checked_instance(ShapeAugmentedGamma, _instance)
new = super(ShapeAugmentedGamma, self).expand(batch_shape, new)
batch_shape = torch.Size(batch_shape)
new.concentration = self.concentration.expand(batch_shape)
new._boost = self._boost
new._rejection_gamma = self._rejection_gamma.expand(batch_shape)
new._validate_args = self._validate_args
return new
def rsample(self, sample_shape=torch.Size()):
x = self._rejection_gamma.rsample(sample_shape)
boosted_x = x.clone()
for i in range(self._boost):
u = torch.rand(x.shape, dtype=x.dtype, device=x.device)
boosted_x *= (1 - u) ** (1 / (i + self.concentration))
self._unboost_x_cache = boosted_x, x
return boosted_x
def score_parts(self, boosted_x=None):
if boosted_x is None:
boosted_x = self._unboost_x_cache[0]
assert boosted_x is self._unboost_x_cache[0]
x = self._unboost_x_cache[1]
_, score_function, _ = self._rejection_gamma.score_parts(x)
log_prob = self.log_prob(boosted_x)
return ScoreParts(log_prob, score_function, log_prob)
@copy_docs_from(Dirichlet)
class ShapeAugmentedDirichlet(Dirichlet):
"""
Implementation of ``Dirichlet`` via ``ShapeAugmentedGamma``.
This naive implementation has stochastic reparameterized gradients, which
have higher variance than PyTorch's ``Dirichlet`` implementation.
"""
def __init__(self, concentration, boost=1, validate_args=None):
super(ShapeAugmentedDirichlet, self).__init__(concentration, validate_args=validate_args)
self._gamma = ShapeAugmentedGamma(concentration, torch.ones_like(concentration), boost)
def expand(self, batch_shape, _instance=None):
new = self._get_checked_instance(ShapeAugmentedDirichlet, _instance)
new = super(ShapeAugmentedDirichlet, self).expand(batch_shape, new)
batch_shape = torch.Size(batch_shape)
new._gamma = self._gamma.expand(batch_shape + self._gamma.concentration.shape[-1:])
new._validate_args = self._validate_args
return new
def rsample(self, sample_shape=torch.Size()):
gammas = self._gamma.rsample(sample_shape)
return gammas / gammas.sum(-1, True)
@copy_docs_from(Beta)
class ShapeAugmentedBeta(Beta):
"""
Implementation of ``rate`` via ``ShapeAugmentedGamma``.
This naive implementation has stochastic reparameterized gradients, which
have higher variance than PyTorch's ``rate`` implementation.
"""
def __init__(self, concentration1, concentration0, boost=1, validate_args=None):
super(ShapeAugmentedBeta, self).__init__(concentration1, concentration0, validate_args=validate_args)
alpha_beta = torch.stack([concentration1, concentration0], -1)
self._gamma = ShapeAugmentedGamma(alpha_beta, torch.ones_like(alpha_beta), boost)
def expand(self, batch_shape, _instance=None):
new = self._get_checked_instance(ShapeAugmentedBeta, _instance)
new = super(ShapeAugmentedBeta, self).expand(batch_shape, new)
batch_shape = torch.Size(batch_shape)
new._gamma = self._gamma.expand(batch_shape + self._gamma.concentration.shape[-1:])
new._validate_args = self._validate_args
return new
def rsample(self, sample_shape=torch.Size()):
gammas = self._gamma.rsample(sample_shape)
probs = gammas / gammas.sum(-1, True)
return probs[..., 0]