{{ message }}
Repository navigation
Expand file tree
/
Copy pathvae_flax.py
More file actions
812 lines (667 loc) · 28.3 KB
/
Copy pathvae_flax.py
File metadata and controls
812 lines (667 loc) · 28.3 KB
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
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
# JAX implementation of VQGAN from taming-transformers https://github.com/CompVis/taming-transformers
import math
from functools import partial
from typing import Tuple
import flax
import flax.linen as nn
import jax
import jax.numpy as jnp
from flax.core.frozen_dict import FrozenDict
from ..configuration_utils import ConfigMixin, flax_register_to_config
from ..modeling_flax_utils import FlaxModelMixin
from ..utils import BaseOutput
@flax.struct.dataclass
class FlaxDecoderOutput(BaseOutput):
"""
Output of decoding method.
Args:
sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)`):
Decoded output sample of the model. Output of the last layer of the model.
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
sample: jnp.ndarray
@flax.struct.dataclass
class FlaxAutoencoderKLOutput(BaseOutput):
"""
Output of AutoencoderKL encoding method.
Args:
latent_dist (`FlaxDiagonalGaussianDistribution`):
Encoded outputs of `Encoder` represented as the mean and logvar of `FlaxDiagonalGaussianDistribution`.
`FlaxDiagonalGaussianDistribution` allows for sampling latents from the distribution.
"""
latent_dist: "FlaxDiagonalGaussianDistribution"
class FlaxUpsample2D(nn.Module):
"""
Flax implementation of 2D Upsample layer
Args:
in_channels (`int`):
Input channels
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int
dtype: jnp.dtype = jnp.float32
def setup(self):
self.conv = nn.Conv(
self.in_channels,
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
def __call__(self, hidden_states):
batch, height, width, channels = hidden_states.shape
hidden_states = jax.image.resize(
hidden_states,
shape=(batch, height * 2, width * 2, channels),
method="nearest",
)
hidden_states = self.conv(hidden_states)
return hidden_states
class FlaxDownsample2D(nn.Module):
"""
Flax implementation of 2D Downsample layer
Args:
in_channels (`int`):
Input channels
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int
dtype: jnp.dtype = jnp.float32
def setup(self):
self.conv = nn.Conv(
self.in_channels,
kernel_size=(3, 3),
strides=(2, 2),
padding="VALID",
dtype=self.dtype,
)
def __call__(self, hidden_states):
pad = ((0, 0), (0, 1), (0, 1), (0, 0)) # pad height and width dim
hidden_states = jnp.pad(hidden_states, pad_width=pad)
hidden_states = self.conv(hidden_states)
return hidden_states
class FlaxResnetBlock2D(nn.Module):
"""
Flax implementation of 2D Resnet Block.
Args:
in_channels (`int`):
Input channels
out_channels (`int`):
Output channels
dropout (:obj:`float`, *optional*, defaults to 0.0):
Dropout rate
use_nin_shortcut (:obj:`bool`, *optional*, defaults to `None`):
Whether to use `nin_shortcut`. This activates a new layer inside ResNet block
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int
out_channels: int = None
dropout: float = 0.0
use_nin_shortcut: bool = None
dtype: jnp.dtype = jnp.float32
def setup(self):
out_channels = self.in_channels if self.out_channels is None else self.out_channels
self.norm1 = nn.GroupNorm(num_groups=32, epsilon=1e-6)
self.conv1 = nn.Conv(
out_channels,
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
self.norm2 = nn.GroupNorm(num_groups=32, epsilon=1e-6)
self.dropout_layer = nn.Dropout(self.dropout)
self.conv2 = nn.Conv(
out_channels,
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
use_nin_shortcut = self.in_channels != out_channels if self.use_nin_shortcut is None else self.use_nin_shortcut
self.conv_shortcut = None
if use_nin_shortcut:
self.conv_shortcut = nn.Conv(
out_channels,
kernel_size=(1, 1),
strides=(1, 1),
padding="VALID",
dtype=self.dtype,
)
def __call__(self, hidden_states, deterministic=True):
residual = hidden_states
hidden_states = self.norm1(hidden_states)
hidden_states = nn.swish(hidden_states)
hidden_states = self.conv1(hidden_states)
hidden_states = self.norm2(hidden_states)
hidden_states = nn.swish(hidden_states)
hidden_states = self.dropout_layer(hidden_states, deterministic)
hidden_states = self.conv2(hidden_states)
if self.conv_shortcut is not None:
residual = self.conv_shortcut(residual)
return hidden_states + residual
class FlaxAttentionBlock(nn.Module):
r"""
Flax Convolutional based multi-head attention block for diffusion-based VAE.
Parameters:
channels (:obj:`int`):
Input channels
num_head_channels (:obj:`int`, *optional*, defaults to `None`):
Number of attention heads
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
channels: int
num_head_channels: int = None
dtype: jnp.dtype = jnp.float32
def setup(self):
self.num_heads = self.channels // self.num_head_channels if self.num_head_channels is not None else 1
dense = partial(nn.Dense, self.channels, dtype=self.dtype)
self.group_norm = nn.GroupNorm(num_groups=32, epsilon=1e-6)
self.query, self.key, self.value = dense(), dense(), dense()
self.proj_attn = dense()
def transpose_for_scores(self, projection):
new_projection_shape = projection.shape[:-1] + (self.num_heads, -1)
# move heads to 2nd position (B, T, H * D) -> (B, T, H, D)
new_projection = projection.reshape(new_projection_shape)
# (B, T, H, D) -> (B, H, T, D)
new_projection = jnp.transpose(new_projection, (0, 2, 1, 3))
return new_projection
def __call__(self, hidden_states):
residual = hidden_states
batch, height, width, channels = hidden_states.shape
hidden_states = self.group_norm(hidden_states)
hidden_states = hidden_states.reshape((batch, height * width, channels))
query = self.query(hidden_states)
key = self.key(hidden_states)
value = self.value(hidden_states)
# transpose
query = self.transpose_for_scores(query)
key = self.transpose_for_scores(key)
value = self.transpose_for_scores(value)
# compute attentions
scale = 1 / math.sqrt(math.sqrt(self.channels / self.num_heads))
attn_weights = jnp.einsum("...qc,...kc->...qk", query * scale, key * scale)
attn_weights = nn.softmax(attn_weights, axis=-1)
# attend to values
hidden_states = jnp.einsum("...kc,...qk->...qc", value, attn_weights)
hidden_states = jnp.transpose(hidden_states, (0, 2, 1, 3))
new_hidden_states_shape = hidden_states.shape[:-2] + (self.channels,)
hidden_states = hidden_states.reshape(new_hidden_states_shape)
hidden_states = self.proj_attn(hidden_states)
hidden_states = hidden_states.reshape((batch, height, width, channels))
hidden_states = hidden_states + residual
return hidden_states
class FlaxDownEncoderBlock2D(nn.Module):
r"""
Flax Resnet blocks-based Encoder block for diffusion-based VAE.
Parameters:
in_channels (:obj:`int`):
Input channels
out_channels (:obj:`int`):
Output channels
dropout (:obj:`float`, *optional*, defaults to 0.0):
Dropout rate
num_layers (:obj:`int`, *optional*, defaults to 1):
Number of Resnet layer block
add_downsample (:obj:`bool`, *optional*, defaults to `True`):
Whether to add downsample layer
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int
out_channels: int
dropout: float = 0.0
num_layers: int = 1
add_downsample: bool = True
dtype: jnp.dtype = jnp.float32
def setup(self):
resnets = []
for i in range(self.num_layers):
in_channels = self.in_channels if i == 0 else self.out_channels
res_block = FlaxResnetBlock2D(
in_channels=in_channels,
out_channels=self.out_channels,
dropout=self.dropout,
dtype=self.dtype,
)
resnets.append(res_block)
self.resnets = resnets
if self.add_downsample:
self.downsamplers_0 = FlaxDownsample2D(self.out_channels, dtype=self.dtype)
def __call__(self, hidden_states, deterministic=True):
for resnet in self.resnets:
hidden_states = resnet(hidden_states, deterministic=deterministic)
if self.add_downsample:
hidden_states = self.downsamplers_0(hidden_states)
return hidden_states
class FlaxUpEncoderBlock2D(nn.Module):
r"""
Flax Resnet blocks-based Encoder block for diffusion-based VAE.
Parameters:
in_channels (:obj:`int`):
Input channels
out_channels (:obj:`int`):
Output channels
dropout (:obj:`float`, *optional*, defaults to 0.0):
Dropout rate
num_layers (:obj:`int`, *optional*, defaults to 1):
Number of Resnet layer block
add_downsample (:obj:`bool`, *optional*, defaults to `True`):
Whether to add downsample layer
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int
out_channels: int
dropout: float = 0.0
num_layers: int = 1
add_upsample: bool = True
dtype: jnp.dtype = jnp.float32
def setup(self):
resnets = []
for i in range(self.num_layers):
in_channels = self.in_channels if i == 0 else self.out_channels
res_block = FlaxResnetBlock2D(
in_channels=in_channels,
out_channels=self.out_channels,
dropout=self.dropout,
dtype=self.dtype,
)
resnets.append(res_block)
self.resnets = resnets
if self.add_upsample:
self.upsamplers_0 = FlaxUpsample2D(self.out_channels, dtype=self.dtype)
def __call__(self, hidden_states, deterministic=True):
for resnet in self.resnets:
hidden_states = resnet(hidden_states, deterministic=deterministic)
if self.add_upsample:
hidden_states = self.upsamplers_0(hidden_states)
return hidden_states
class FlaxUNetMidBlock2D(nn.Module):
r"""
Flax Unet Mid-Block module.
Parameters:
in_channels (:obj:`int`):
Input channels
dropout (:obj:`float`, *optional*, defaults to 0.0):
Dropout rate
num_layers (:obj:`int`, *optional*, defaults to 1):
Number of Resnet layer block
attn_num_head_channels (:obj:`int`, *optional*, defaults to `1`):
Number of attention heads for each attention block
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int
dropout: float = 0.0
num_layers: int = 1
attn_num_head_channels: int = 1
dtype: jnp.dtype = jnp.float32
def setup(self):
# there is always at least one resnet
resnets = [
FlaxResnetBlock2D(
in_channels=self.in_channels,
out_channels=self.in_channels,
dropout=self.dropout,
dtype=self.dtype,
)
]
attentions = []
for _ in range(self.num_layers):
attn_block = FlaxAttentionBlock(
channels=self.in_channels, num_head_channels=self.attn_num_head_channels, dtype=self.dtype
)
attentions.append(attn_block)
res_block = FlaxResnetBlock2D(
in_channels=self.in_channels,
out_channels=self.in_channels,
dropout=self.dropout,
dtype=self.dtype,
)
resnets.append(res_block)
self.resnets = resnets
self.attentions = attentions
def __call__(self, hidden_states, deterministic=True):
hidden_states = self.resnets[0](hidden_states, deterministic=deterministic)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
hidden_states = attn(hidden_states)
hidden_states = resnet(hidden_states, deterministic=deterministic)
return hidden_states
class FlaxEncoder(nn.Module):
r"""
Flax Implementation of VAE Encoder.
This model is a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
general usage and behavior.
Finally, this model supports inherent JAX features such as:
- [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
- [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
- [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
- [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
Parameters:
in_channels (:obj:`int`, *optional*, defaults to 3):
Input channels
out_channels (:obj:`int`, *optional*, defaults to 3):
Output channels
down_block_types (:obj:`Tuple[str]`, *optional*, defaults to `(DownEncoderBlock2D)`):
DownEncoder block type
block_out_channels (:obj:`Tuple[str]`, *optional*, defaults to `(64,)`):
Tuple containing the number of output channels for each block
layers_per_block (:obj:`int`, *optional*, defaults to `2`):
Number of Resnet layer for each block
norm_num_groups (:obj:`int`, *optional*, defaults to `2`):
norm num group
act_fn (:obj:`str`, *optional*, defaults to `silu`):
Activation function
double_z (:obj:`bool`, *optional*, defaults to `False`):
Whether to double the last output channels
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
Parameters `dtype`
"""
in_channels: int = 3
out_channels: int = 3
down_block_types: Tuple[str] = ("DownEncoderBlock2D",)
block_out_channels: Tuple[int] = (64,)
layers_per_block: int = 2
norm_num_groups: int = 32
act_fn: str = "silu"
double_z: bool = False
dtype: jnp.dtype = jnp.float32
def setup(self):
block_out_channels = self.block_out_channels
# in
self.conv_in = nn.Conv(
block_out_channels[0],
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
# downsampling
down_blocks = []
output_channel = block_out_channels[0]
for i, _ in enumerate(self.down_block_types):
input_channel = output_channel
output_channel = block_out_channels[i]
is_final_block = i == len(block_out_channels) - 1
down_block = FlaxDownEncoderBlock2D(
in_channels=input_channel,
out_channels=output_channel,
num_layers=self.layers_per_block,
add_downsample=not is_final_block,
dtype=self.dtype,
)
down_blocks.append(down_block)
self.down_blocks = down_blocks
# middle
self.mid_block = FlaxUNetMidBlock2D(
in_channels=block_out_channels[-1], attn_num_head_channels=None, dtype=self.dtype
)
# end
conv_out_channels = 2 * self.out_channels if self.double_z else self.out_channels
self.conv_norm_out = nn.GroupNorm(num_groups=32, epsilon=1e-6)
self.conv_out = nn.Conv(
conv_out_channels,
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
def __call__(self, sample, deterministic: bool = True):
# in
sample = self.conv_in(sample)
# downsampling
for block in self.down_blocks:
sample = block(sample, deterministic=deterministic)
# middle
sample = self.mid_block(sample, deterministic=deterministic)
# end
sample = self.conv_norm_out(sample)
sample = nn.swish(sample)
sample = self.conv_out(sample)
return sample
class FlaxDecoder(nn.Module):
r"""
Flax Implementation of VAE Decoder.
This model is a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
general usage and behavior.
Finally, this model supports inherent JAX features such as:
- [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
- [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
- [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
- [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
Parameters:
in_channels (:obj:`int`, *optional*, defaults to 3):
Input channels
out_channels (:obj:`int`, *optional*, defaults to 3):
Output channels
up_block_types (:obj:`Tuple[str]`, *optional*, defaults to `(UpDecoderBlock2D)`):
UpDecoder block type
block_out_channels (:obj:`Tuple[str]`, *optional*, defaults to `(64,)`):
Tuple containing the number of output channels for each block
layers_per_block (:obj:`int`, *optional*, defaults to `2`):
Number of Resnet layer for each block
norm_num_groups (:obj:`int`, *optional*, defaults to `32`):
norm num group
act_fn (:obj:`str`, *optional*, defaults to `silu`):
Activation function
double_z (:obj:`bool`, *optional*, defaults to `False`):
Whether to double the last output channels
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
parameters `dtype`
"""
in_channels: int = 3
out_channels: int = 3
up_block_types: Tuple[str] = ("UpDecoderBlock2D",)
block_out_channels: int = (64,)
layers_per_block: int = 2
norm_num_groups: int = 32
act_fn: str = "silu"
dtype: jnp.dtype = jnp.float32
def setup(self):
block_out_channels = self.block_out_channels
# z to block_in
self.conv_in = nn.Conv(
block_out_channels[-1],
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
# middle
self.mid_block = FlaxUNetMidBlock2D(
in_channels=block_out_channels[-1], attn_num_head_channels=None, dtype=self.dtype
)
# upsampling
reversed_block_out_channels = list(reversed(block_out_channels))
output_channel = reversed_block_out_channels[0]
up_blocks = []
for i, _ in enumerate(self.up_block_types):
prev_output_channel = output_channel
output_channel = reversed_block_out_channels[i]
is_final_block = i == len(block_out_channels) - 1
up_block = FlaxUpEncoderBlock2D(
in_channels=prev_output_channel,
out_channels=output_channel,
num_layers=self.layers_per_block + 1,
add_upsample=not is_final_block,
dtype=self.dtype,
)
up_blocks.append(up_block)
prev_output_channel = output_channel
self.up_blocks = up_blocks
# end
self.conv_norm_out = nn.GroupNorm(num_groups=32, epsilon=1e-6)
self.conv_out = nn.Conv(
self.out_channels,
kernel_size=(3, 3),
strides=(1, 1),
padding=((1, 1), (1, 1)),
dtype=self.dtype,
)
def __call__(self, sample, deterministic: bool = True):
# z to block_in
sample = self.conv_in(sample)
# middle
sample = self.mid_block(sample, deterministic=deterministic)
# upsampling
for block in self.up_blocks:
sample = block(sample, deterministic=deterministic)
sample = self.conv_norm_out(sample)
sample = nn.swish(sample)
sample = self.conv_out(sample)
return sample
class FlaxDiagonalGaussianDistribution(object):
def __init__(self, parameters, deterministic=False):
# Last axis to account for channels-last
self.mean, self.logvar = jnp.split(parameters, 2, axis=-1)
self.logvar = jnp.clip(self.logvar, -30.0, 20.0)
self.deterministic = deterministic
self.std = jnp.exp(0.5 * self.logvar)
self.var = jnp.exp(self.logvar)
if self.deterministic:
self.var = self.std = jnp.zeros_like(self.mean)
def sample(self, key):
return self.mean + self.std * jax.random.normal(key, self.mean.shape)
def kl(self, other=None):
if self.deterministic:
return jnp.array([0.0])
if other is None:
return 0.5 * jnp.sum(self.mean**2 + self.var - 1.0 - self.logvar, axis=[1, 2, 3])
return 0.5 * jnp.sum(
jnp.square(self.mean - other.mean) / other.var + self.var / other.var - 1.0 - self.logvar + other.logvar,
axis=[1, 2, 3],
)
def nll(self, sample, axis=[1, 2, 3]):
if self.deterministic:
return jnp.array([0.0])
logtwopi = jnp.log(2.0 * jnp.pi)
return 0.5 * jnp.sum(logtwopi + self.logvar + jnp.square(sample - self.mean) / self.var, axis=axis)
def mode(self):
return self.mean
@flax_register_to_config
class FlaxAutoencoderKL(nn.Module, FlaxModelMixin, ConfigMixin):
r"""
Flax Implementation of Variational Autoencoder (VAE) model with KL loss from the paper Auto-Encoding Variational
Bayes by Diederik P. Kingma and Max Welling.
This model is a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
general usage and behavior.
Finally, this model supports inherent JAX features such as:
- [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
- [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
- [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
- [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
Parameters:
in_channels (:obj:`int`, *optional*, defaults to 3):
Input channels
out_channels (:obj:`int`, *optional*, defaults to 3):
Output channels
down_block_types (:obj:`Tuple[str]`, *optional*, defaults to `(DownEncoderBlock2D)`):
DownEncoder block type
up_block_types (:obj:`Tuple[str]`, *optional*, defaults to `(UpDecoderBlock2D)`):
UpDecoder block type
block_out_channels (:obj:`Tuple[str]`, *optional*, defaults to `(64,)`):
Tuple containing the number of output channels for each block
layers_per_block (:obj:`int`, *optional*, defaults to `2`):
Number of Resnet layer for each block
act_fn (:obj:`str`, *optional*, defaults to `silu`):
Activation function
latent_channels (:obj:`int`, *optional*, defaults to `4`):
Latent space channels
norm_num_groups (:obj:`int`, *optional*, defaults to `32`):
Norm num group
sample_size (:obj:`int`, *optional*, defaults to `32`):
Sample input size
dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
parameters `dtype`
"""
in_channels: int = 3
out_channels: int = 3
down_block_types: Tuple[str] = ("DownEncoderBlock2D",)
up_block_types: Tuple[str] = ("UpDecoderBlock2D",)
block_out_channels: Tuple[int] = (64,)
layers_per_block: int = 1
act_fn: str = "silu"
latent_channels: int = 4
norm_num_groups: int = 32
sample_size: int = 32
dtype: jnp.dtype = jnp.float32
def setup(self):
self.encoder = FlaxEncoder(
in_channels=self.config.in_channels,
out_channels=self.config.latent_channels,
down_block_types=self.config.down_block_types,
block_out_channels=self.config.block_out_channels,
layers_per_block=self.config.layers_per_block,
act_fn=self.config.act_fn,
norm_num_groups=self.config.norm_num_groups,
double_z=True,
dtype=self.dtype,
)
self.decoder = FlaxDecoder(
in_channels=self.config.latent_channels,
out_channels=self.config.out_channels,
up_block_types=self.config.up_block_types,
block_out_channels=self.config.block_out_channels,
layers_per_block=self.config.layers_per_block,
norm_num_groups=self.config.norm_num_groups,
act_fn=self.config.act_fn,
dtype=self.dtype,
)
self.quant_conv = nn.Conv(
2 * self.config.latent_channels,
kernel_size=(1, 1),
strides=(1, 1),
padding="VALID",
dtype=self.dtype,
)
self.post_quant_conv = nn.Conv(
self.config.latent_channels,
kernel_size=(1, 1),
strides=(1, 1),
padding="VALID",
dtype=self.dtype,
)
def init_weights(self, rng: jax.random.PRNGKey) -> FrozenDict:
# init input tensors
sample_shape = (1, self.in_channels, self.sample_size, self.sample_size)
sample = jnp.zeros(sample_shape, dtype=jnp.float32)
params_rng, dropout_rng, gaussian_rng = jax.random.split(rng, 3)
rngs = {"params": params_rng, "dropout": dropout_rng, "gaussian": gaussian_rng}
return self.init(rngs, sample)["params"]
def encode(self, sample, deterministic: bool = True, return_dict: bool = True):
sample = jnp.transpose(sample, (0, 2, 3, 1))
hidden_states = self.encoder(sample, deterministic=deterministic)
moments = self.quant_conv(hidden_states)
posterior = FlaxDiagonalGaussianDistribution(moments)
if not return_dict:
return (posterior,)
return FlaxAutoencoderKLOutput(latent_dist=posterior)
def decode(self, latents, deterministic: bool = True, return_dict: bool = True):
if latents.shape[-1] != self.config.latent_channels:
latents = jnp.transpose(latents, (0, 2, 3, 1))
hidden_states = self.post_quant_conv(latents)
hidden_states = self.decoder(hidden_states, deterministic=deterministic)
hidden_states = jnp.transpose(hidden_states, (0, 3, 1, 2))
if not return_dict:
return (hidden_states,)
return FlaxDecoderOutput(sample=hidden_states)
def __call__(self, sample, sample_posterior=False, deterministic: bool = True, return_dict: bool = True):
posterior = self.encode(sample, deterministic=deterministic, return_dict=return_dict)
if sample_posterior:
rng = self.make_rng("gaussian")
hidden_states = posterior.latent_dist.sample(rng)
else:
hidden_states = posterior.latent_dist.mode()
sample = self.decode(hidden_states, return_dict=return_dict).sample
if not return_dict:
return (sample,)
return FlaxDecoderOutput(sample=sample)
You can’t perform that action at this time.