-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathvariational_autoencoder.py
638 lines (575 loc) · 22.1 KB
/
variational_autoencoder.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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
from __future__ import absolute_import, division, print_function
import inspect
from collections import defaultdict
from dataclasses import dataclass, field
from functools import partial
from itertools import zip_longest
from typing import (Any, Callable, Dict, Iterator, List, Optional, Tuple, Type,
Union, Sequence)
from typing_extensions import Literal
import numpy as np
import tensorflow as tf
from six import string_types
from tensorflow import Tensor
from tensorflow.python import keras
from tensorflow.python.keras.layers import Layer
from tensorflow_probability.python import layers as tfl
from tensorflow_probability.python.distributions import Distribution
from tqdm import tqdm
from odin import backend as bk
from odin.backend import TensorType
from odin.backend.keras_helpers import layer2text
from odin.bay.layers.dense_distribution import DistributionDense
from odin.bay.random_variable import RVconf
from odin.bay.vi._base import VariationalModel
from odin.networks import Identity, NetConf, TrainStep, SequentialNetwork
from odin.utils import as_tuple
from odin.bay.vi.utils import traverse_dims
__all__ = [
'LayerCreator',
'VAEStep',
'VariationalAutoencoder',
'SemiSupervisedVAE',
'VAE',
]
# ===========================================================================
# Types
# ===========================================================================
LayerCreator = Union[str, Layer, Type[Layer], NetConf, RVconf,
Callable[[Optional[List[int]]], Layer]]
# ===========================================================================
# Helpers
# ===========================================================================
def _get_args(layer):
spec = inspect.getfullargspec(layer.call)
return set(spec.args + spec.kwonlyargs)
def _net2str(net):
if isinstance(net, (keras.Sequential, tfl.DistributionLambda)):
return layer2text(net)
return str(net)
def _parse_layers(network, name=None
) -> Union[Layer, DistributionDense, SequentialNetwork]:
## make sure is a list
if isinstance(network, (tuple, list)):
if len(network) != 1:
raise ValueError(
f'Only support single neural network but provide {network}')
network = network[0]
assert network is not None, 'network cannot be None'
## check different options
cfg = network
## identity
if cfg is None:
layer = Identity(name=name)
## string alias of activation
elif isinstance(cfg, string_types):
layer = keras.layers.Activation(keras.activations.get(cfg))
## Callable or type
elif (inspect.isfunction(cfg) or isinstance(cfg, partial) or
isinstance(cfg, type)):
layer = cfg()
## RVmeta
elif isinstance(cfg, RVconf):
layer = cfg.create_posterior(name=name if cfg.name is None else None)
## the NetConf
elif isinstance(cfg, NetConf):
layer = cfg.create_network(name=name)
## Layer
elif isinstance(cfg, Layer):
layer = cfg
## no support
else:
raise ValueError(
f"No support for network configuration of type: {type(cfg)}")
return layer
def _iter_lists(X, Y):
r""" Try to match the length of list-Y to list-X,
the yield a pair of (x, y) with the condition x is not None """
Y = Y * len(X) if len(Y) == 1 else Y
for i, (x, y) in enumerate(zip_longest(X, Y)):
if x is not None:
yield i, x, y
# ===========================================================================
# Training step
# ===========================================================================
@dataclass
class VAEStep(TrainStep):
""" A single train step (iteration) for Variational Autoencoder """
vae: 'VariationalAutoencoder' = None
call_kw: Dict[str, Any] = field(default_factory=dict)
def call(self) -> Tuple[Tensor, Dict[str, Any]]:
llk, kl = self.vae.elbo_components(self.inputs,
training=self.training,
mask=self.mask,
**self.call_kw)
# check if array is empty, return 0s
loss = -tf.reduce_mean(self.vae.elbo(llk, kl))
metrics = dict(**{k: tf.reduce_mean(v) for k, v in llk.items()},
**{k: tf.reduce_mean(v) for k, v in kl.items()})
return loss, metrics
# ===========================================================================
# Model
# ===========================================================================
class VariationalAutoencoder(VariationalModel):
"""Base class for all variational autoencoder
Parameters
----------
encoder : LayerCreator, optional
the encoder network, by default NetConf()
decoder : LayerCreator, optional
the decoder network, by default NetConf()
outputs : LayerCreator, optional
a descriptor for the input/output, by default
`RVmeta(64, 'gaus', projection=True, name="Input")`
latents : LayerCreator, optional
a descriptor for the latents' distribution, by default
`RVmeta(10, 'mvndiag', projection=True, name="Latent")`
input_shape : Optional[List[int]], optional
specific input_shape for the network, if not given, use the given `outputs`,
by default None
Returns
---------
p(X|Z) : a single or a list of `tensorflow_probability.Distribution`
q(Z|X) : a single or a list of `tensorflow_probability.Distribution`
Layers
--------
encoder : list of `keras.layers.Layer`.
Encoding inputs to latents
decoder : list of `keras.layers.Layer`.
Decoding latents to intermediate states
latents : list of `keras.layers.Layer`.
A list of the Dense layer that create the latent variable (random variable)
observation : list of `keras.layers.Layer`.
A list of the Dense layer that create the output variable
(random or deterministic variable)
"""
def __init__(
self,
observation: LayerCreator = RVconf((28, 28, 1), 'bernoulli',
projection=True,
name='image'),
latents: Optional[LayerCreator] = RVconf(16, 'mvndiag',
projection=True,
name="latents"),
encoder: Optional[LayerCreator] = None,
decoder: Optional[LayerCreator] = None,
**kwargs,
):
if encoder is None:
encoder = NetConf((512, 512), flatten_inputs=True, name="encoder")
if decoder is None:
decoder = NetConf((512, 512), flatten_inputs=True, flatten_outputs=True,
name="decoder")
### keras want this supports_masking on to enable support masking
super().__init__(**kwargs)
### create layers
# encoder
if isinstance(encoder, (tuple, list)):
self._encoder = [
_parse_layers(network=e, name=f"encoder{i}")
for i, e in enumerate(encoder)
]
self._encoder_args = [_get_args(e) for e in self._encoder]
else:
self._encoder = _parse_layers(network=encoder, name="encoder")
self._encoder_args = _get_args(self.encoder)
# latents
if isinstance(latents, (tuple, list)):
self._latents = [
_parse_layers(network=z, name=f"latents{i}")
for i, z in enumerate(latents)
]
self._latents_args = [_get_args(z) for z in self.latents]
else:
self._latents = _parse_layers(network=latents, name="latents")
self._latents_args = _get_args(self.latents)
# decoder
if isinstance(decoder, (tuple, list)):
self._decoder = [
_parse_layers(network=d, name=f"decoder{i}")
for i, d in enumerate(decoder)
]
self._decoder_args = [_get_args(d) for d in self.decoder]
else:
self._decoder = _parse_layers(network=decoder, name="decoder")
self._decoder_args = _get_args(self.decoder)
# observation
if isinstance(observation, (tuple, list)):
self._observation = [
_parse_layers(network=observation, name=f"observation{i}")
for i, o in enumerate(observation)
]
self._observation_args = [_get_args(o) for o in self.observation]
else:
self._observation = _parse_layers(network=observation, name="observation")
self._observation_args = _get_args(self.observation)
@property
def encoder(self) -> Union[SequentialNetwork, List[SequentialNetwork]]:
return self._encoder
@property
def decoder(self) -> Union[SequentialNetwork, List[SequentialNetwork]]:
return self._decoder
@property
def latents(self) -> Union[DistributionDense, List[DistributionDense]]:
return self._latents
@property
def observation(self) -> Union[DistributionDense, List[DistributionDense]]:
return self._observation
@property
def n_latents(self) -> int:
if isinstance(self.latents, (tuple, list)):
return len(self.latents)
return 1
@property
def n_observation(self) -> int:
if isinstance(self.observation, (tuple, list)):
return len(self.observation)
return 1
@property
def input_shape(self) -> Sequence[Union[None, int]]:
return self.encoder.input_shape
@property
def latent_shape(self) -> Sequence[Union[None, int]]:
return self.decoder.input_shape
def sample_prior(self, n: int = 1, seed: int = 1, **kwargs) -> Tensor:
"""Sampling from prior distribution"""
return bk.atleast_2d(
self.latents.sample(sample_shape=n, seed=seed))
def sample_observation_prior(self, n: int = 1, seed: int = 1) -> Tensor:
"""Sample from observation prior, i.e. p(X), given that the prior of X
is known.
Notes
-----
This could be wrong since `RVmeta` often has a default prior.
"""
return bk.atleast_2d(
self.observation.sample(sample_shape=n, seed=seed))
def sample_observation(self, n: int = 1, seed: int = 1,
training: bool = False, **kwargs) -> Distribution:
"""Sample observation using latents prior"""
z = self.sample_prior(n, seed=seed)
return self.decode(z, training=training)
def encode(self,
inputs: Union[TensorType, List[TensorType]],
training: Optional[bool] = None,
mask: Optional[TensorType] = None,
only_encoding: bool = False,
**kwargs) -> Distribution:
""" Encoding inputs to latent codes """
kw = dict(kwargs)
if 'mask' in self._encoder_args:
kw['mask'] = mask
if 'training' in self._encoder_args:
kw['training'] = training
h_e = self.encoder(inputs, **kw)
if only_encoding:
return h_e
# create the latents distribution
kw = {}
if 'training' in self._latents_args:
kw['training'] = training
if 'mask' in self._latents_args:
kw['mask'] = mask
if 'sample_shape' in self._latents_args:
kw['sample_shape'] = self.sample_shape
qz_x = self.latents(h_e, **kw)
# need to keep the keras mask
qz_x._keras_mask = mask
return qz_x
def decode(self,
latents: Union[TensorType, List[TensorType]],
training: Optional[bool] = None,
mask: Optional[Tensor] = None,
only_decoding: bool = False,
**kwargs) -> Distribution:
""" Decoding latent codes, this does not guarantee output the
reconstructed distribution """
# stop tensorflow complaining about tensor inputs for Sequential
c = tf.constant(0., dtype=self.dtype)
if isinstance(latents, (tuple, list)):
latents = [qz + c for qz in latents]
else:
latents = latents + c
# flatten the sample shapes
if self.sample_ndim > 0:
ndim = (latents.shape.ndims - self.sample_ndim - 1)
flat_shape = tf.shape(latents)[-ndim:]
flat_shape = tf.concat([[-1], flat_shape], axis=0)
latents = tf.reshape(latents, flat_shape)
# apply the decoder and get back the sample shape
kw = dict(kwargs)
if 'mask' in self._decoder_args:
kw['mask'] = mask
if 'training' in self._decoder_args:
kw['training'] = training
h_d = self.decoder(latents, **kw)
# recover the sample shape
if self.sample_ndim > 0:
org_shape = tf.concat(
[self.sample_shape, [-1], tf.shape(h_d)[1:]], axis=0)
h_d = tf.reshape(h_d, org_shape)
# only return hidden states from the decoder
if only_decoding:
return h_d
# create the output distribution
kw = {}
if 'training' in self._observation_args:
kw['training'] = training
if 'mask' in self._observation_args:
kw['mask'] = mask
px_z = self.observation(h_d, **kw)
# remember to store the keras mask in outputs
px_z._keras_mask = mask
return px_z
def call(self,
inputs: TensorType,
training: Optional[bool] = None,
mask: Optional[Tensor] = None,
**kwargs) -> Tuple[Distribution, Distribution]:
"""Applying the encode-decode process for VAE
Parameters
----------
inputs : TensorTypes
inputs' Tensors
training : Optional[bool], optional
training or evaluation mode, by default None
mask : Optional[Tensor], optional
mask, by default None
Returns
-------
Union[Distribution, List[Distribution]]
`p_{theta}(x||z)` the output distribution(s)
Union[Distribution, List[Distribution]]
`q_{\phi}(z||x)` the latent distribution(s)
"""
# encode
qz_x = self.encode(inputs, training=training, mask=mask, **kwargs)
# transfer the mask from encoder to decoder here
for qz in as_tuple(qz_x):
if hasattr(qz, '_keras_mask') and qz._keras_mask is not None:
mask = qz._keras_mask
break
# decode
px_z = self.decode(qz_x, training=training, mask=mask, **kwargs)
return px_z, qz_x
def marginal_log_prob(
self,
inputs: Union[TensorType, List[TensorType]],
training: Optional[bool] = None,
n_mcmc: Optional[int] = 100,
reduce: Optional[Callable[[Tensor], Tensor]] = tf.reduce_mean,
batch_size: int = 32,
verbose: bool = False,
**kwargs,
) -> Tuple[Dict[str, Tensor], Dict[str, Tensor]]:
"""Marginal log likelihood `log(p(X))`, an biased estimation.
With sufficient amount of MCMC samples (-> inf), the value will converges
to `log(p(X))`
With large amount of sample, recommending reduce the batch size to very
small number or use CPU for the calculation `with tf.device("/CPU:0"):`
Note: this function will need further modification for more complicated
prior and latent space, only work for:
- vanilla-VAE or
- with proper prior injected into qZ_X and pZ_X using
`qZ_X.KL_divergence.prior = ...` during `encode` or `decode` methods
Parameters
----------
inputs : TensorTypes
inputs' Tensors
training : Optional[bool], optional
training or evaluation mode, by default None
Returns
-------
Tuple[Dict[str, Tensor], Dict[str, Tensor]]
reconstruction : mapping from name to marginal reconstruction
KL divergence: mapping from name to the divergence
"""
## prepare data
if isinstance(inputs, (tuple, list)):
pass
elif not isinstance(inputs, tf.data.Dataset):
inputs = tf.data.Dataset.from_tensor_slices(inputs).batch(int(batch_size))
## check the MCMC shape
if n_mcmc is None:
n_mcmc = int(np.prod(self.sample_shape))
else:
n_mcmc = int(n_mcmc)
## single step
@tf.function
def _step(X):
ret_llk = []
ret_kl = []
if isinstance(X, dict):
Q = self.encode(training=training, **X, **kwargs)
else:
Q = self.encode(X, training=training, **kwargs)
Q = as_tuple(Q)
z = [i.sample(n_mcmc) for i in Q]
z_reshape = [tf.reshape(i, (-1, i.shape[-1])) for i in z]
P = self.decode(z_reshape[0] if len(Q) == 1 else z_reshape,
training=training)
P = as_tuple(P)
# calculate the KL
for qz, z in zip(Q, z):
if hasattr(qz, 'KL_divergence'):
pz = qz.KL_divergence.prior
name = qz.name.split('_')[0]
llk_q = qz.log_prob(z)
llk_p = pz.log_prob(z)
ret_kl.append((name, (llk_q, llk_p)))
# calculate the LLK
if isinstance(X, dict):
X = X['inputs']
X = as_tuple(X)
for px, x in zip(P, X):
x = tf.tile(x, [n_mcmc] + [1 for i in range(len(x.shape) - 1)])
name = px.name.split('_')[0]
llk_x = tf.reshape(px.log_prob(x), (n_mcmc, -1))
ret_llk.append((name, llk_x))
return ret_llk, ret_kl
## run the inputs
llk = defaultdict(list)
kl = defaultdict(list)
old_sample_shape = self.sample_shape
self._sample_shape = ()
inputs = tqdm(inputs, desc='MarginalLLK', disable=not verbose)
for n_batch, X in enumerate(inputs):
ret_llk, ret_kl = _step(X)
for name, val in ret_llk:
llk[str(name.numpy(), 'utf-8')].append(val)
for name, val in ret_kl:
kl[str(name.numpy(), 'utf-8')].append(val)
self._sample_shape = old_sample_shape
inputs.clear()
inputs.close()
# concatenate
C = tf.math.log(tf.cast(n_mcmc, self.dtype))
def logsumexp_concat(x):
return tf.reduce_logsumexp(tf.concat(x, axis=-1), axis=0)
llk = {
name: logsumexp_concat(logprobs) - C for name, logprobs in llk.items()
}
kl = {
name: (
logsumexp_concat([i for i, _ in llkqp]) - C,
logsumexp_concat([i for _, i in llkqp]) - C,
) for name, llkqp in kl.items()
}
## Marginal LLK
if reduce is not None:
llk = {i: reduce(j) for i, j in llk.items()}
kl = {i: (reduce(q), reduce(p)) for i, (q, p) in kl.items()}
return llk, kl
def elbo_components(
self,
inputs: Union[TensorType, List[TensorType]],
training: Optional[bool] = None,
mask: Optional[Tensor] = None,
**kwargs,
) -> Tuple[Dict[str, Tensor], Dict[str, Tensor]]:
"""Calculate the distortion (log-likelihood) and rate (KL-divergence)
for contruction the Evident Lower Bound (ELBO)"""
# organize all inputs to list
px_z, qz_x = self(inputs, training=training, mask=mask, **kwargs)
### llk
llk = {}
for o, x, px in zip(as_tuple(self.observation), as_tuple(inputs),
as_tuple(px_z)):
llk[f'llk_{o.name}'] = px.log_prob(x)
### kl
kl = {}
for z, qz in zip(as_tuple(self.latents), as_tuple(qz_x)):
if hasattr(qz, "KL_divergence"):
kl[f'kl_{z.name}'] = qz.KL_divergence(analytic=self.analytic,
reverse=self.reverse,
free_bits=self.free_bits,
sample_shape=None,
keepdims=True)
else:
kl[f'kl_{z.name}'] = tf.constant(0., dtype=self.dtype)
return llk, kl
################## For training
def train_steps(self,
inputs: Union[TensorType, List[TensorType]],
training: Optional[bool] = None,
mask: Optional[Tensor] = None,
name: str = '',
**kwargs) -> Iterator[VAEStep]:
"""Support multiple steps training for each iteration (similar to GAN)"""
yield VAEStep(vae=self,
parameters=self.trainable_variables,
inputs=inputs,
training=training,
mask=mask,
name=name,
call_kw=kwargs)
def __str__(self):
cls = [
i for i in type.mro(type(self)) if issubclass(i, VariationalAutoencoder)
]
text = (f"{'->'.join([i.__name__ for i in cls[::-1]])} "
f"(semi:{type(self).is_semi_supervised()})")
text += f'\n Tensorboard : {self.tensorboard_logdir}'
text += f'\n Analytic : {self.analytic}'
text += f'\n Reverse : {self.reverse}'
text += f'\n Free-bits : {self.free_bits}'
text += f'\n Sample Shape : {self.sample_shape}'
text += f'\n Fitted : {int(self.step.numpy())}(iters)'
text += f'\n MD5 checksum : {self.md5_checksum}'
## encoder
for i, encoder in enumerate(as_tuple(self.encoder)):
text += f"\n Encoder#{i}:\n "
text += "\n ".join(_net2str(encoder).split('\n'))
## Decoder
for i, decoder in enumerate(as_tuple(self.decoder)):
text += f"\n Decoder#{i}:\n "
text += "\n ".join(_net2str(decoder).split('\n'))
## Latent
for i, latent in enumerate(as_tuple(self.latents)):
text += f"\n Latent#{i}:\n "
text += "\n ".join(_net2str(latent).split('\n'))
## Output
for i, output in enumerate(as_tuple(self.observation)):
text += f"\n Output#{i}:\n "
text += "\n ".join(_net2str(output).split('\n'))
## Optimizer
if hasattr(self, 'optimizer'):
for i, opt in enumerate(as_tuple(self.optimizer)):
if isinstance(opt, tf.optimizers.Optimizer):
text += f"\n Optimizer#{i}:\n "
text += "\n ".join(
["%s:%s" % (k, str(v)) for k, v in opt.get_config().items()])
return text
# ===========================================================================
# Semi-supervised VAE
# ===========================================================================
def _is_empty_zero():
return 0.
class SemiSupervisedVAE(VariationalAutoencoder):
""" NOTE: this layer only 1 inputs and multi-labels"""
@classmethod
def is_semi_supervised(cls) -> bool:
return True
def ignore_empty(self,
is_empty: Tensor,
loss_dict: Dict[str, Tensor]
) -> Dict[str, Tensor]:
return {k: tf.cond(is_empty,
true_fn=_is_empty_zero,
false_fn=lambda: v)
for k, v in loss_dict.items()}
def merge_objectives(self,
llk_uns: Dict[str, Tensor],
kl_uns: Dict[str, Tensor],
llk_sup: Dict[str, Tensor],
kl_sup: Dict[str, Tensor],
) -> Tuple[Dict[str, Tensor], Dict[str, Tensor]]:
llk = dict(**{f'uns/{k}': v for k, v in llk_uns.items()},
**{f'sup/{k}': tf.reduce_mean(v) for k, v in llk_sup.items()})
kl = dict(**{f'uns/{k}': v for k, v in kl_uns.items()},
**{f'sup/{k}': tf.reduce_mean(v) for k, v in kl_sup.items()})
return llk, kl
# ===========================================================================
# Others
# ===========================================================================
VAE = VariationalAutoencoder