SCAE_Attack/model.py at master · FrostbiteXSW/SCAE_Attack · GitHub

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
import sonnet as snt
import tensorflow as tf

from capsules import primary
from capsules.attention import SetTransformer
from capsules.models.scae import ImageAutoencoder
from capsules.models.scae import ImageCapsule


def stacked_capsule_autoencoders(
		canvas_size,
		template_size=11,
		n_part_caps=16,
		n_part_caps_dims=6,
		n_part_special_features=16,
		part_encoder_noise_scale=0.,
		n_channels=1,
		colorize_templates=False,
		use_alpha_channel=False,
		template_nonlin='relu1',
		color_nonlin='relu1',
		n_obj_caps=10,
		n_obj_caps_params=32,
		obj_decoder_noise_type=None,
		obj_decoder_noise_scale=0.,
		num_classes=10,
		prior_within_example_sparsity_weight=1.,
		prior_between_example_sparsity_weight=1.,
		posterior_within_example_sparsity_weight=10.,
		posterior_between_example_sparsity_weight=10.,
		set_transformer_n_layers=3,
		set_transformer_n_heads=1,
		set_transformer_n_dims=16,
		set_transformer_n_output_dims=256,
		part_cnn_strides=None,
		prep='none',
		scope='stacked_capsule_autoencoders'
):
	if part_cnn_strides is None:
		part_cnn_strides = [2, 2, 1, 1]

	"""Builds the SCAE."""
	with tf.variable_scope(scope, 'stacked_capsule_autoencoders'):
		img_size = [canvas_size] * 2
		template_size = [template_size] * 2

		cnn_encoder = snt.nets.ConvNet2D(
			output_channels=[128] * 4,
			kernel_shapes=[3],
			strides=part_cnn_strides,
			paddings=[snt.VALID],
			activate_final=True
		)

		part_encoder = primary.CapsuleImageEncoder(
			cnn_encoder,
			n_part_caps,
			n_part_caps_dims,
			n_features=n_part_special_features,
			similarity_transform=False,
			encoder_type='conv_att',
			noise_scale=part_encoder_noise_scale
		)

		part_decoder = primary.TemplateBasedImageDecoder(
			output_size=img_size,
			template_size=template_size,
			n_channels=n_channels,
			learn_output_scale=False,
			colorize_templates=colorize_templates,
			use_alpha_channel=use_alpha_channel,
			template_nonlin=template_nonlin,
			color_nonlin=color_nonlin,
		)

		obj_encoder = SetTransformer(
			n_layers=set_transformer_n_layers,
			n_heads=set_transformer_n_heads,
			n_dims=set_transformer_n_dims,
			n_output_dims=set_transformer_n_output_dims,
			n_outputs=n_obj_caps,
			layer_norm=True,
			dropout_rate=0.)

		obj_decoder = ImageCapsule(
			n_obj_caps,
			2,
			n_part_caps,
			n_caps_params=n_obj_caps_params,
			n_hiddens=128,
			learn_vote_scale=True,
			deformations=True,
			noise_type=obj_decoder_noise_type,
			noise_scale=obj_decoder_noise_scale,
			similarity_transform=False
		)

		model = ImageAutoencoder(
			primary_encoder=part_encoder,
			primary_decoder=part_decoder,
			encoder=obj_encoder,
			decoder=obj_decoder,
			input_key='image',
			label_key='label',
			n_classes=num_classes,
			dynamic_l2_weight=10,
			caps_ll_weight=1.,
			vote_type='enc',
			pres_type='enc',
			stop_grad_caps_inpt=True,
			stop_grad_caps_target=True,
			prior_sparsity_loss_type='l2',
			prior_within_example_sparsity_weight=prior_within_example_sparsity_weight,
			prior_between_example_sparsity_weight=prior_between_example_sparsity_weight,
			posterior_sparsity_loss_type='entropy',
			posterior_within_example_sparsity_weight=posterior_within_example_sparsity_weight,
			posterior_between_example_sparsity_weight=posterior_between_example_sparsity_weight,
			prep=prep
		)

	return model