/
layers.py
262 lines (214 loc) · 7.72 KB
/
layers.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
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
# Copyright 2019 The TensorFlow Probability Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Base classes for building neural networks."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import tensorflow.compat.v2 as tf
from tensorflow_probability.python.experimental.nn import util as nn_util_lib
from tensorflow_probability.python.internal import name_util
__all__ = [
'Lambda',
'Layer',
'Sequential',
]
class Layer(tf.Module):
"""A `callable` `tf.Module` characterized by `eval(input)`."""
def __init__(self, also_track=None, name=None):
name = name_util.strip_invalid_chars(name or type(self).__name__)
self._also_track = [] if also_track is None else [also_track]
super(Layer, self).__init__(name=name)
self._extra_loss = None
self._extra_result = None
self._trace = False
@property
def extra_loss(self):
return self._extra_loss
@property
def extra_result(self):
return self._extra_result
@property
def also_track(self):
return list(self._also_track)
def eval(self, inputs, is_training=True, **kwargs):
self._set_extra_loss(None)
self._set_extra_result(None)
return inputs
def summary(self):
return nn_util_lib.variables_summary(self.variables, self.name)
def save(self, filename):
return nn_util_lib.variables_save(filename, self.variables)
def load(self, filename):
return nn_util_lib.variables_load(filename, self.variables)
def __call__(self, inputs, **kwargs):
if callable(inputs):
return Sequential([inputs, self], **kwargs)
self._extra_loss = self._extra_result = None
y = self.eval(inputs, **kwargs)
# TODO(jvdillon): Consider adding provenance.
# y.__tfp_nn_provenance = self
return y
def __repr__(self):
return '<{}: name={}>'.format(type(self).__name__, self.name)
def _set_extra_result(self, value):
self._extra_result = value
def _set_extra_loss(self, value):
self._extra_loss = value
class Sequential(Layer):
"""A `Layer` characterized by iteratively given functions."""
def __init__(self, layers, also_track=None, name=None):
layers = tuple(layers)
if not layers:
raise tf.errors.InvalidArgumentError(
'Argument `layers` must contain at least one element.')
name = name or '_'.join([_try_get_name(x) for x in layers])
self._layers = tuple(layers)
super(Sequential, self).__init__(also_track=also_track, name=name)
def set_trace(self, trace):
self._trace = bool(trace)
return self
@property
def layers(self):
return self._layers
def eval(self, inputs, is_training=True, **kwargs):
kwargs.update({'is_training': is_training})
all_extras = []
x = inputs
if self._trace:
_trace(self, x, -1)
for i, layer in enumerate(self.layers):
_try_set_extra_results(layer, loss=None, result=None)
x = _try_call(layer, [x], kwargs)
if self._trace:
_trace(layer, x, i)
extra_loss, extra_result = _try_get_extra_results(layer)
all_extras.append((extra_loss, extra_result))
_try_set_extra_results(layer, loss=extra_loss, result=extra_result)
non_none_extra_losses = [extra_loss for (extra_loss, _) in all_extras
if extra_loss is not None]
sum_extra_losses = (sum(non_none_extra_losses)
if non_none_extra_losses else None)
self._set_extra_loss(sum_extra_losses)
self._set_extra_result(None)
return x
def __getitem__(self, i):
r = Sequential(self.layers[i], name=self.name)
r._also_track = self._also_track # pylint: disable=protected-access
return r
class Lambda(Layer):
"""A `Layer` which can be defined inline."""
def __init__(self,
eval_fn=None,
extra_loss_fn=None,
also_track=None,
name=None):
if not callable(eval_fn):
raise tf.errors.InvalidArgumentError(
'Argument `eval_fn` must be `callable`.')
name = name or _try_get_name(eval_fn)
self._eval_fn = eval_fn
self._extra_loss_fn = extra_loss_fn
super(Lambda, self).__init__(also_track=also_track, name=name)
def eval(self, inputs, is_training=True, **kwargs):
kwargs.update({'is_training': is_training})
if self._eval_fn is not None:
r = _try_call(self._eval_fn, [inputs], kwargs)
else:
r = inputs
self._last_call = r # For variable tracking purposes.
self._set_extra_loss(None if self._extra_loss_fn is None else
_try_call(self._extra_loss_fn, [r], kwargs))
self._set_extra_result(None)
return r
class KernelBiasLayer(Layer):
"""Linear layer."""
def __init__(self,
kernel,
bias,
apply_kernel_fn,
activation_fn=None,
dtype=tf.float32,
name=None):
self._kernel = kernel
self._bias = bias
self._activation_fn = activation_fn
self._apply_kernel_fn = apply_kernel_fn
self._dtype = dtype
super(KernelBiasLayer, self).__init__(name=name)
@property
def dtype(self):
return self._dtype
@property
def kernel(self):
return self._kernel
@property
def bias(self):
return self._bias
@property
def activation_fn(self):
return self._activation_fn
def eval(self, x, is_training=True):
x = tf.convert_to_tensor(x, dtype_hint=self.dtype, name='x')
y = x
if self.kernel is not None:
y = self._apply_kernel_fn(y, self.kernel)
if self.bias is not None:
y = y + self.bias
if self.activation_fn is not None:
y = self.activation_fn(y) # pylint: disable=not-callable
return y
def _try_set_extra_results(layer, loss, result):
"""Convenience function for maybe calling `_set_extra_result`."""
set_fn = getattr(layer, '_set_extra_loss', None)
if callable(set_fn):
set_fn(loss)
set_fn = getattr(layer, '_set_extra_result', None)
if callable(set_fn):
set_fn(result)
def _try_get_extra_results(layer):
"""Convenience function for getting side data."""
return (
getattr(layer, 'extra_loss', None),
getattr(layer, 'extra_result', None),
)
def _try_call(fn, args, kwargs):
"""Convenience function for evaluating argument `fn`."""
try:
if fn is None:
return args[0]
try:
return fn(*args, **kwargs)
except TypeError:
# Don't return from here or else we'll pick up a nested exception.
# Seeing TypeError here isn't really an exception since it only means we
# need to call `fn` differently).
pass
return fn(*args)
except:
print('------ EXCEPTION in {} ------'.format(_try_get_name(fn)))
raise
def _try_get_name(fn, name_fallback='unknown'):
return str(getattr(fn, '__name__', None) or
getattr(fn, 'name', None) or
getattr(type(fn), '__name__', name_fallback))
def _trace(layer, x, i):
name = _try_get_name(layer)
z = tf.nest.map_structure(lambda x_: '{:14} {:<24} {:>10}'.format( # pylint: disable=g-long-lambda
_try_get_name(x_),
str(list(getattr(x_, 'shape', '?'))),
_try_get_name(getattr(x_, 'dtype', x_), '?')), x)
print('--- TRACE{:02}: {:<24} {}'.format(i, name, z))
sys.stdout.flush()