forked from google-deepmind/open_x_embodiment
-
Notifications
You must be signed in to change notification settings - Fork 0
/
efficientnet.py
643 lines (535 loc) · 18.4 KB
/
efficientnet.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
639
640
641
642
643
"""Defines variants of the EfficientNet model.
Currently implements the follow variants:
- EfficientNetWithFilm - EfficientNet backbone with FiLM-conditioning applied.
"""
import copy
import math
from typing import Any, Optional, Sequence, Tuple, Union
from flax import linen as nn
from flax.linen import initializers
import jax
from jax import numpy as jnp
import film_conditioning
MEAN_RGB = [0.485, 0.456, 0.406]
STDDEV_RGB = [0.229, 0.224, 0.225]
# Relevant initializers. The original implementation uses fan_out Kaiming init.
conv_kernel_init_fn = initializers.variance_scaling(2.0, 'fan_out', 'normal')
dense_kernel_init_fn = initializers.variance_scaling(
1 / 3.0, 'fan_out', 'uniform'
)
class DepthwiseConv(nn.Module):
"""Depthwise convolution that matches tensorflow's conventions.
In Tensorflow, the shapes of depthwise kernels don't match the shapes of a
regular convolutional kernel of appropriate feature_group_count.
It is safer to use this class instead of the regular Conv (easier port of
tensorflow checkpoints, fan_out initialization of the previous layer will
match the tensorflow behavior, etc...).
Attributes:
features: Number of convolution filters.
kernel_size: Shape of the convolutional kernel.
strides: A sequence of `n` integers, representing the inter-window strides.
padding: Either the string `'SAME'`, the string `'VALID'`, or a sequence of
`n` `(low, high)` integer pairs that give the padding to apply before and
after each spatial dimension.
input_dilation: `None`, or a sequence of `n` integers, giving the dilation
factor to apply in each spatial dimension of `inputs`. Convolution with
input dilation `d` is equivalent to transposed convolution with stride
`d`.
kernel_dilation: `None`, or a sequence of `n` integers, giving the dilation
factor to apply in each spatial dimension of the convolution kernel.
Convolution with kernel dilation is also known as 'atrous convolution'.
feature_group_count: Unused attribute present in nn.Conv. Declare it to
match the nn.Conv API.
use_bias: Whether to add a bias to the output (default: True).
dtype: The dtype of the computation (default: float32).
precision: Numerical precision of the computation see `jax.lax.Precision`
for details.
kernel_init: Initializer for the convolutional kernel.
bias_init: Initializer for the bias.
"""
features: int
kernel_size: Tuple[int, int]
strides: Optional[Tuple[int, int]] = None
padding: Union[str, Sequence[int]] = 'SAME'
input_dilation: Optional[Sequence[int]] = None
kernel_dilation: Optional[Sequence[int]] = None
feature_group_count: int = 1
use_bias: bool = True
dtype: jnp.dtype = jnp.float32
precision: Any = None
kernel_init: Any = initializers.lecun_normal()
bias_init: Any = initializers.zeros
@nn.compact
def __call__(self, inputs: jnp.ndarray) -> jnp.ndarray:
"""Applies a convolution to the inputs.
Args:
inputs: Input data with dimensions (batch, spatial_dims..., features).
Returns:
The convolved data.
"""
inputs = jnp.asarray(inputs, self.dtype)
in_features = inputs.shape[-1]
strides = self.strides
if strides is None:
strides = (1,) * (inputs.ndim - 2)
kernel_shape = self.kernel_size + (self.features, 1)
# Naming convention follows tensorflow.
kernel = self.param('depthwise_kernel', self.kernel_init, kernel_shape)
kernel = jnp.asarray(kernel, self.dtype)
# Need to transpose to convert tensorflow-shaped kernel to lax-shaped kernel
kernel = jnp.transpose(kernel, [0, 1, 3, 2])
dimension_numbers = nn.linear._conv_dimension_numbers(inputs.shape) # pylint:disable=protected-access
y = jax.lax.conv_general_dilated(
inputs,
kernel,
strides,
self.padding,
lhs_dilation=self.input_dilation,
rhs_dilation=self.kernel_dilation,
dimension_numbers=dimension_numbers,
feature_group_count=in_features,
precision=self.precision,
)
if self.use_bias:
bias = self.param('bias', self.bias_init, (self.features,))
bias = jnp.asarray(bias, self.dtype)
y = y + bias
return y
# pytype: disable=attribute-error
# pylint:disable=unused-argument
class BlockConfig(object):
"""Class that contains configuration parameters for a single block."""
def __init__(
self,
input_filters: int = 0,
output_filters: int = 0,
kernel_size: int = 3,
num_repeat: int = 1,
expand_ratio: int = 1,
strides: Tuple[int, int] = (1, 1),
se_ratio: Optional[float] = None,
id_skip: bool = True,
fused_conv: bool = False,
conv_type: str = 'depthwise',
):
for arg in locals().items():
setattr(self, *arg)
class ModelConfig(object):
"""Class that contains configuration parameters for the model."""
def __init__(
self,
width_coefficient: float = 1.0,
depth_coefficient: float = 1.0,
resolution: int = 224,
dropout_rate: float = 0.2,
blocks: Tuple[BlockConfig, ...] = (
# (input_filters, output_filters, kernel_size, num_repeat,
# expand_ratio, strides, se_ratio)
# pylint: disable=bad-whitespace
BlockConfig(32, 16, 3, 1, 1, (1, 1), 0.25),
BlockConfig(16, 24, 3, 2, 6, (2, 2), 0.25),
BlockConfig(24, 40, 5, 2, 6, (2, 2), 0.25),
BlockConfig(40, 80, 3, 3, 6, (2, 2), 0.25),
BlockConfig(80, 112, 5, 3, 6, (1, 1), 0.25),
BlockConfig(112, 192, 5, 4, 6, (2, 2), 0.25),
BlockConfig(192, 320, 3, 1, 6, (1, 1), 0.25),
# pylint: enable=bad-whitespace
),
stem_base_filters: int = 32,
top_base_filters: int = 1280,
activation: str = 'swish',
batch_norm: str = 'default',
bn_momentum: float = 0.99,
bn_epsilon: float = 1e-3,
# While the original implementation used a weight decay of 1e-5,
# tf.nn.l2_loss divides it by 2, so we halve this to compensate in Keras
weight_decay: float = 5e-6,
drop_connect_rate: float = 0.2,
depth_divisor: int = 8,
min_depth: Optional[int] = None,
use_se: bool = True,
input_channels: int = 3,
num_classes: int = 1000,
model_name: str = 'efficientnet',
rescale_input: bool = True,
data_format: str = 'channels_last',
final_projection_size: int = 0,
classifier_head: bool = True,
dtype: jnp.dtype = jnp.float32,
):
"""Default Config for Efficientnet-B0."""
for arg in locals().items():
setattr(self, *arg)
# pylint:enable=unused-argument
MODEL_CONFIGS = {
# (width, depth, resolution, dropout)
'efficientnet-b3': ModelConfig(1.2, 1.4, 300, 0.3),
}
def round_filters(filters: int, config: ModelConfig) -> int:
"""Returns rounded number of filters based on width coefficient."""
width_coefficient = config.width_coefficient
min_depth = config.min_depth
divisor = config.depth_divisor
if not width_coefficient:
return filters
filters *= width_coefficient
min_depth = min_depth or divisor
new_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_filters < 0.9 * filters:
new_filters += divisor
return int(new_filters)
def round_repeats(repeats: int, depth_coefficient: float) -> int:
"""Returns rounded number of repeats based on depth coefficient."""
return int(math.ceil(depth_coefficient * repeats))
def conv2d(
inputs: jnp.ndarray,
num_filters: int,
config: ModelConfig,
kernel_size: Union[int, Tuple[int, int]] = (1, 1),
strides: Tuple[int, int] = (1, 1),
use_batch_norm: bool = True,
use_bias: bool = False,
activation: Any = None,
depthwise: bool = False,
train: bool = True,
conv_name: Optional[str] = None,
bn_name: Optional[str] = None,
dtype: jnp.dtype = jnp.float32,
) -> jnp.ndarray:
"""Convolutional layer with possibly batch norm and activation.
Args:
inputs: Input data with dimensions (batch, spatial_dims..., features).
num_filters: Number of convolution filters.
config: Configuration for the model.
kernel_size: Size of the kernel, as a tuple of int.
strides: Strides for the convolution, as a tuple of int.
use_batch_norm: Whether batch norm should be applied to the output.
use_bias: Whether we should add bias to the output of the first convolution.
activation: Name of the activation function to use.
depthwise: If true, will use depthwise convolutions.
train: Whether the model should behave in training or inference mode.
conv_name: Name to give to the convolution layer.
bn_name: Name to give to the batch norm layer.
dtype: dtype for the computation.
Returns:
The output of the convolutional layer.
"""
conv_fn = DepthwiseConv if depthwise else nn.Conv
kernel_size = (
(kernel_size, kernel_size)
if isinstance(kernel_size, int)
else tuple(kernel_size)
)
conv_name = conv_name if conv_name else 'conv2d'
bn_name = bn_name if bn_name else 'batch_normalization'
x = conv_fn(
num_filters,
kernel_size,
tuple(strides),
padding='SAME',
use_bias=use_bias,
kernel_init=conv_kernel_init_fn,
name=conv_name,
dtype=dtype,
)(inputs)
if use_batch_norm:
x = nn.BatchNorm(
use_running_average=not train,
momentum=config.bn_momentum,
epsilon=config.bn_epsilon,
name=bn_name,
dtype=dtype,
)(x)
if activation is not None:
x = getattr(nn.activation, activation.lower())(x)
return x
def stochastic_depth(
inputs: jnp.ndarray,
rng: jnp.ndarray,
survival_probability: float,
deterministic: bool = False,
) -> jnp.ndarray:
"""Applies stochastic depth.
Args:
inputs: The inputs that should be randomly masked.
rng: A `jax.random.PRNGKey`.
survival_probability: 1 - the probability of masking out a value.
deterministic: If false the inputs are scaled by `1 / (1 - rate)` and
masked, whereas if true, no mask is applied and the inputs are returned as
is.
Returns:
The masked inputs.
"""
if survival_probability == 1.0 or deterministic:
return inputs
mask_shape = [inputs.shape[0]] + [1 for _ in inputs.shape[1:]]
mask = jax.random.bernoulli(rng, p=survival_probability, shape=mask_shape)
mask = jnp.tile(mask, [1] + list(inputs.shape[1:]))
return jax.lax.select(
mask, inputs / survival_probability, jnp.zeros_like(inputs)
)
class SqueezeExcite(nn.Module):
"""SqueezeExite block (See: https://arxiv.org/abs/1709.01507.)
Attributes:
num_filters: Number of convolution filters.
block: Configuration for this block.
config: Configuration for the model.
train: Whether the model is in training or inference mode.
"""
num_filters: int
block: BlockConfig
config: ModelConfig
train: bool
@nn.compact
def __call__(self, inputs: jnp.ndarray) -> jnp.ndarray:
"""Applies a convolution to the inputs.
Args:
inputs: Input data with dimensions (batch, spatial_dims..., features).
Returns:
The output of the squeeze excite block.
"""
block = self.block
config = self.config
train = self.train
dtype = config.dtype
num_reduced_filters = max(1, int(block.input_filters * block.se_ratio))
se = nn.avg_pool(inputs, inputs.shape[1:3])
se = conv2d(
se,
num_reduced_filters,
config,
use_bias=True,
use_batch_norm=False,
activation=config.activation,
conv_name='reduce_conv2d_0',
train=train,
dtype=dtype,
)
se = conv2d(
se,
self.num_filters,
config,
use_bias=True,
use_batch_norm=False,
activation='sigmoid',
conv_name='expand_conv2d_0',
train=train,
dtype=dtype,
)
return inputs * se
class MBConvBlock(nn.Module):
"""Main building component of Efficientnet.
Attributes:
block: BlockConfig, arguments to create a Block.
config: ModelConfig, a set of model parameters.
train: Whether we are training or predicting.
"""
block: BlockConfig
config: ModelConfig
train: bool = False
@nn.compact
def __call__(self, inputs: jnp.ndarray) -> jnp.ndarray:
"""Mobile Inverted Residual Bottleneck.
Args:
inputs: Input to the block.
Returns:
The output of the block.
"""
config = self.config
block = self.block
train = self.train
use_se = config.use_se
activation = config.activation
drop_connect_rate = config.drop_connect_rate
use_depthwise = block.conv_type != 'no_depthwise'
dtype = config.dtype
rng = self.make_rng('random')
filters = block.input_filters * block.expand_ratio
x = inputs
bn_index = 0
if block.fused_conv:
# If we use fused mbconv, skip expansion and use regular conv.
x = conv2d(
x,
filters,
config,
kernel_size=block.kernel_size,
strides=block.strides,
activation=activation,
conv_name='fused_conv2d_0',
bn_name='batch_normalization_' + str(bn_index),
train=train,
dtype=dtype,
)
bn_index += 1
else:
if block.expand_ratio != 1:
# Expansion phase
kernel_size = (1, 1) if use_depthwise else (3, 3)
x = conv2d(
x,
filters,
config,
kernel_size=kernel_size,
activation=activation,
conv_name='expand_conv2d_0',
bn_name='batch_normalization_' + str(bn_index),
train=train,
dtype=dtype,
)
bn_index += 1
# Depthwise Convolution
if use_depthwise:
x = conv2d(
x,
num_filters=x.shape[-1], # Depthwise conv
config=config,
kernel_size=block.kernel_size,
strides=block.strides,
activation=activation,
depthwise=True,
conv_name='depthwise_conv2d',
bn_name='batch_normalization_' + str(bn_index),
train=train,
dtype=dtype,
)
bn_index += 1
# Squeeze and Excitation phase
if use_se:
assert block.se_ratio is not None
assert 0 < block.se_ratio <= 1
x = SqueezeExcite(
num_filters=filters, block=block, config=config, train=train
)(x)
# Output phase
x = conv2d(
x,
block.output_filters,
config,
activation=None,
conv_name='project_conv2d_0',
bn_name='batch_normalization_' + str(bn_index),
train=train,
dtype=dtype,
)
if (
block.id_skip
and all(s == 1 for s in block.strides)
and block.input_filters == block.output_filters
):
if drop_connect_rate and drop_connect_rate > 0:
survival_probability = 1 - drop_connect_rate
x = stochastic_depth(
x, rng, survival_probability, deterministic=not train
)
x = x + inputs
return x
class Stem(nn.Module):
"""Initial block of Efficientnet.
Attributes:
config: ModelConfig, a set of model parameters.
train: Whether we are training or predicting.
"""
config: ModelConfig
train: bool = False
@nn.compact
def __call__(self, inputs: jnp.ndarray) -> jnp.ndarray:
"""Returns the output of the stem block.
Args:
inputs: The input to the block.
Returns:
Output of the block
"""
config = self.config
train = self.train
x = conv2d(
inputs,
round_filters(config.stem_base_filters, config),
config,
kernel_size=(3, 3),
strides=(2, 2),
activation=config.activation,
train=train,
dtype=config.dtype,
)
return x
class Head(nn.Module):
"""Final block of Efficientnet.
Attributes:
config: A set of model parameters.
train: Whether we are training or predicting.
"""
config: Any
train: bool = True
@nn.compact
def __call__(self, inputs: jnp.ndarray) -> jnp.ndarray:
"""Returns the output of the head block.
Args:
inputs: The input to the block.
Returns:
x: Classifier logits.
"""
config = self.config
train = self.train
dtype = config.dtype
# Build top.
x = conv2d(
inputs,
round_filters(config.top_base_filters, config),
config,
activation=config.activation,
train=train,
dtype=dtype,
)
return x
# pytype: enable=attribute-error
class EfficientNetWithFilm(nn.Module):
"""EfficientNet with FiLM conditioning."""
config: Any
dtype: jnp.dtype = jnp.float32
@nn.compact
def __call__(
self, inputs: jnp.ndarray, context_input: jnp.ndarray, *, train: bool
):
"""Returns the output of the EfficientNet model."""
config = copy.deepcopy(self.config)
config.dtype = self.dtype
depth_coefficient = config.depth_coefficient
blocks = config.blocks
drop_connect_rate = config.drop_connect_rate
inputs = jnp.asarray(inputs, self.dtype)
# Build stem.
x = Stem(config=config, train=train)(inputs)
# Build blocks.
num_blocks_total = sum(
round_repeats(block.num_repeat, depth_coefficient) for block in blocks
)
block_num = 0
for _, block in enumerate(blocks):
assert block.num_repeat > 0
# Update block input and output filters based on depth multiplier.
block.input_filters = round_filters(block.input_filters, config)
block.output_filters = round_filters(block.output_filters, config)
block.num_repeat = round_repeats(block.num_repeat, depth_coefficient)
# The first block needs to take care of stride and filter size increase
drop_rate = drop_connect_rate * float(block_num) / num_blocks_total
config.drop_connect_rate = drop_rate
x = MBConvBlock(block=block, config=config, train=train)(x)
x = film_conditioning.FilmConditioning(num_channels=x.shape[-1])(
x, context_input
)
block_num += 1
if block.num_repeat > 1:
block.input_filters = block.output_filters
block.strides = [1, 1]
for _ in range(block.num_repeat - 1):
drop_rate = drop_connect_rate * float(block_num) / num_blocks_total
config.drop_connect_rate = drop_rate
x = MBConvBlock(block=block, config=config, train=train)(x)
x = film_conditioning.FilmConditioning(num_channels=x.shape[-1])(
x, context_input
)
block_num += 1
x = Head(self.config, train=train)(x)
return x