/
irm_vae.py
181 lines (161 loc) · 6.51 KB
/
irm_vae.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
from typing import Optional
import numpy as np
import tensorflow as tf
from odin.bay.layers import DistributionDense, VectorDeterministicLayer
from odin.bay.vi.autoencoder.variational_autoencoder import (
LayerCreator, RVconf, VariationalAutoencoder, _iter_lists)
from odin.networks import SequentialNetwork
from odin.utils import as_tuple
from tensorflow.python import keras
from tensorflow.python.eager import context
from tensorflow.python.training.tracking import base as trackable
class ImplicitRankMinimizer(keras.layers.Layer):
"""Implicit rank-minimization (Variational) Autoencoder
The idea is to add a link-list of linear weights between the encoder and the
bottleneck, since stochastic gradient descent going to force a low-rank solution
on the latent codes.
Parameters
----------
units : int
number of latent units
n_layers : int, optional
number of linear weights to be added, by default 3
share_weights : bool, optional
all layers share the same weight matrix, by default False
References
---------
Jing, L., Zbontar, J. & LeCun, Y. Implicit Rank-Minimizing Autoencoder.
arXiv:2010.00679 [cs, stat] (2020).
"""
def __init__(self,
units: int,
n_layers: int = 3,
share_weights: bool = False,
initializer: str = 'glorot_uniform',
regularizer: Optional[str] = None,
constraint: Optional[str] = None,
activity_regularizer: Optional[str] = None,
name: str = 'IRM'):
super().__init__(activity_regularizer=activity_regularizer, name=name)
self.units = int(units)
self.n_layers = int(n_layers)
self.share_weights = bool(share_weights)
self.linear_weights = []
self.initializer = as_tuple(keras.initializers.get(initializer), N=n_layers)
self.regularizer = as_tuple(keras.regularizers.get(regularizer), N=n_layers)
self.constraint = as_tuple(keras.constraints.get(constraint), N=n_layers)
def build(self, input_shape) -> 'ImplicitRankMinimizer':
super().build(input_shape)
input_dim = input_shape[-1]
last_w = None
for i, (init, regu, cons) in enumerate(
zip(self.initializer, self.regularizer, self.constraint)):
shape = (input_dim, self.units)
if self.share_weights and last_w is not None and last_w.shape == shape:
w = last_w
else:
w = self.add_weight(name=f'W{i}',
shape=shape,
dtype=self.dtype,
initializer=init,
regularizer=regu,
constraint=cons,
trainable=True)
last_w = w
input_dim = w.shape[-1]
self.linear_weights.append(w)
return self
def call(self, inputs, training=None, **kwargs):
for kernel in self.linear_weights:
rank = inputs.shape.rank
if rank == 2 or rank is None:
if isinstance(inputs, tf.sparse.SparseTensor):
inputs = tf.sparse.sparse_dense_matmul(inputs, kernel)
else:
inputs = tf.matmul(inputs, kernel)
# Broadcast kernel to inputs.
else:
shape = inputs.shape.as_list()
inputs = tf.tensordot(inputs, kernel, [[rank - 1], [0]])
# Reshape the output back to the original ndim of the input.
if not context.executing_eagerly():
output_shape = shape[:-1] + [kernel.shape[-1]]
inputs.set_shape(output_shape)
return inputs
def __str__(self):
return (f'<IRM units:{self.units} n_layers:{self.n_layers} '
f'share_weights:{self.share_weights}>')
class irmVAE(VariationalAutoencoder):
"""Implicit rank-minimization (Variational) Autoencoder
The idea is to add a link-list of linear weights between the encoder and the
bottleneck, since stochastic gradient descent going to force a low-rank solution
on the latent codes.
Parameters
----------
n_layers : int, optional
number of linear weights to be added, by default 3
share_weights : bool, optional
all layers share the same weight matrix, by default False
References
---------
Jing, L., Zbontar, J. & LeCun, Y. Implicit Rank-Minimizing Autoencoder.
arXiv:2010.00679 [cs, stat] (2020).
"""
def __init__(self,
latents: LayerCreator = RVconf(64,
'mvndiag',
projection=True,
name='Latents'),
n_layers: int = 3,
share_weights: bool = False,
name: str = 'irmVAE',
**kwargs):
super().__init__(latents=latents, name=name, **kwargs)
new_encoder = []
for i, layer in enumerate(self.encoder):
# with trackable.no_automatic_dependency_tracking_scope(self):
layer_list = [
layer,
ImplicitRankMinimizer(units=64,
n_layers=n_layers,
share_weights=share_weights,
name='IRM')
]
layer = SequentialNetwork(
layers=layer_list,
name=f'{layer.name}_irm',
)
new_encoder.append(layer)
self._encoder = new_encoder
class irmAE(irmVAE):
def __init__(self,
latents: LayerCreator = RVconf(64,
'vdeterministic',
projection=True,
name='Latents'),
name: str = 'irmAE',
**kwargs):
for qz in tf.nest.flatten(latents):
if isinstance(qz, RVconf):
qz.posterior = 'vdeterministic'
elif isinstance(qz, DistributionDense):
assert qz.posterior == VectorDeterministicLayer, \
('irmAE only support VectorDeterministic posterior, '
f'but given:{qz.posterior}')
super().__init__(latents=latents, name=name, **kwargs)
def elbo_components(self,
inputs,
training=None,
pX_Z=None,
qZ_X=None,
mask=None,
**kwargs):
llk, kl = super().elbo_components(inputs=inputs,
training=training,
pX_Z=pX_Z,
qZ_X=qZ_X,
mask=mask,
**kwargs)
# this make sure no KL is leaking
kl = {k: 0. for k, v in kl.items()}
return llk, kl