TensorFlow-Advanced-Segmentation-Models/tensorflow_advanced_segmentation_models/models/HRNetOCR.py at master · JanMarcelKezmann/TensorFlow-Advanced-Segmentation-Models · 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
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
import tensorflow as tf
import tensorflow.keras.backend as K

from ._custom_layers_and_blocks import ConvolutionBnActivation, BottleneckBlock, HighResolutionModule,SpatialGather_Module, SpatialOCR_Module
from ..backbones.tf_backbones import create_base_model

################################################################################
# High Resolution Network + Object-Contextual Representations
################################################################################
class HRNetOCR(tf.keras.Model):
    def __init__(self, n_classes, filters=64, height=None, width=None, final_activation="softmax",
                 spatial_ocr_scale=1, spatial_context_scale=1, **kwargs):
        super(HRNetOCR, self).__init__(**kwargs)

        self.n_classes = n_classes
        self.filters = filters
        self.height = height
        self.width = width
        self.final_activation = final_activation
        self.spatial_ocr_scale = spatial_ocr_scale
        self.spatial_context_scale = spatial_context_scale

        axis = 3 if K.image_data_format() == "channels_last" else 1

        # Stem Net
        ### Probably set strides to default, i.e. (1, 1)
        self.conv3x3_bn_relu_1 = ConvolutionBnActivation(filters, (3, 3))
        self.conv3x3_bn_relu_2 = ConvolutionBnActivation(filters, (3, 3))

        # stage 1
        self.bottleneck_downsample = BottleneckBlock(64, downsample=True)
        self.bottleneck_1 = BottleneckBlock(64)
        self.bottleneck_2 = BottleneckBlock(64)
        self.bottleneck_3 = BottleneckBlock(64)

        # Stage 2
        # Transition
        self.conv3x3_bn_relu_stage2_1 = ConvolutionBnActivation(48, (3, 3), momentum=0.1)
        self.conv3x3_bn_relu_stage2_2 = ConvolutionBnActivation(96, (3, 3), strides=(2, 2), momentum=0.1)

        # Stage
        # num_modules=1, num_branches=2, blocks=[4, 4], channels=[48, 96]
        self.hrn_stage2_module_1 = HighResolutionModule(num_branches=2, blocks=[4, 4], filters=[48, 96])
        self.hrn_stage2_module_2 = HighResolutionModule(num_branches=2, blocks=[4, 4], filters=[48, 96])

        # Stage 3
        # Transition
        self.conv3x3_bn_relu_stage3 = ConvolutionBnActivation(192, (3, 3), strides=(2, 2), momentum=0.1)

        # Stage
        # num_modules=4, num_branches=3, blocks=[4, 4, 4], channels=[48, 96, 192]
        self.hrn_stage3_module_1 = HighResolutionModule(num_branches=3, blocks=[4, 4, 4], filters=[48, 96, 192])
        self.hrn_stage3_module_2 = HighResolutionModule(num_branches=3, blocks=[4, 4, 4], filters=[48, 96, 192])
        self.hrn_stage3_module_3 = HighResolutionModule(num_branches=3, blocks=[4, 4, 4], filters=[48, 96, 192])
        self.hrn_stage3_module_4 = HighResolutionModule(num_branches=3, blocks=[4, 4, 4], filters=[48, 96, 192])

        # Stage 4
        # Transition
        self.conv3x3_bn_relu_stage4 = ConvolutionBnActivation(384, (3, 3), strides=(2, 2), momentum=0.1)

        # Stage
        # num_modules=3, num_branches=4, num_blocks=[4, 4, 4, 4], num_channels=[48, 96, 192, 384]
        self.hrn_stage4_module_1 = HighResolutionModule(num_branches=4, blocks=[4, 4, 4, 4], filters=[48, 96, 192, 384])
        self.hrn_stage4_module_2 = HighResolutionModule(num_branches=4, blocks=[4, 4, 4, 4], filters=[48, 96, 192, 384])
        self.hrn_stage4_module_3 = HighResolutionModule(num_branches=4, blocks=[4, 4, 4, 4], filters=[48, 96, 192, 384])

        # Upsampling and Concatentation of stages
        self.upsample_x2 = tf.keras.layers.UpSampling2D(size=2, interpolation="bilinear")
        self.upsample_x4 = tf.keras.layers.UpSampling2D(size=4, interpolation="bilinear")
        self.upsample_x8 = tf.keras.layers.UpSampling2D(size=8, interpolation="bilinear")

        self.concat = tf.keras.layers.Concatenate(axis=axis)

        # OCR
        self.aux_head = tf.keras.Sequential([
            ConvolutionBnActivation(720, (1, 1)),
            tf.keras.layers.Conv2D(filters=self.n_classes, kernel_size=(1, 1), use_bias=True),
            tf.keras.layers.Activation(final_activation)
            ])
        self.conv3x3_bn_relu_ocr = ConvolutionBnActivation(512, (3, 3))

        self.spatial_context = SpatialGather_Module(scale=spatial_context_scale)
        self.spatial_ocr = SpatialOCR_Module(512, scale=spatial_ocr_scale, dropout=0.05)

        self.final_conv3x3 = tf.keras.layers.Conv2D(filters=self.n_classes, kernel_size=(1, 1), use_bias=True)
        self.final_activation = tf.keras.layers.Activation(final_activation)

    def call(self, inputs, training=None, mask=None):
        if training is None:
            training = True

        x = self.conv3x3_bn_relu_1(inputs, training=training)
        x = self.conv3x3_bn_relu_2(x, training=training)

        # Stage 1
        x = self.bottleneck_downsample(x, training=training)
        x = self.bottleneck_1(x, training=training)
        x = self.bottleneck_2(x, training=training)
        x = self.bottleneck_3(x, training=training)

        # Stage 2
        x_1 = self.conv3x3_bn_relu_stage2_1(x, training=training)
        x_2 = self.conv3x3_bn_relu_stage2_2(x, training=training) # includes strided convolution

        y_list = self.hrn_stage2_module_1(x_1, x_2, None, None, training=training)
        y_list = self.hrn_stage2_module_2(y_list[0], y_list[1], None, None, training=training)

        # Stage 3
        x_3 = self.conv3x3_bn_relu_stage3(y_list[1], training=training) # includes strided convolution

        y_list = self.hrn_stage3_module_1(y_list[0], y_list[1], x_3, None, training=training)
        y_list = self.hrn_stage3_module_2(y_list[0], y_list[1], y_list[2], None, training=training)
        y_list = self.hrn_stage3_module_3(y_list[0], y_list[1], y_list[2], None, training=training)
        y_list = self.hrn_stage3_module_4(y_list[0], y_list[1], y_list[2], None, training=training)

        # Stage 4
        x_4 = self.conv3x3_bn_relu_stage4(y_list[2], training=training)

        y_list = self.hrn_stage4_module_1(y_list[0], y_list[1], y_list[2], x_4, training=training)
        y_list = self.hrn_stage4_module_2(y_list[0], y_list[1], y_list[2], y_list[3], training=training)
        y_list = self.hrn_stage4_module_3(y_list[0], y_list[1], y_list[2], y_list[3], training=training)

        # Upsampling + Concatentation
        x_2 = self.upsample_x2(y_list[1])
        x_3 = self.upsample_x4(y_list[2])
        x_4 = self.upsample_x8(y_list[3])

        feats = self.concat([y_list[0], x_2, x_3, x_4])

        # OCR
        aux = self.aux_head(feats)

        feats = self.conv3x3_bn_relu_ocr(feats)

        context = self.spatial_context(feats, aux)
        feats = self.spatial_ocr(feats, context)

        out = self.final_conv3x3(feats)
        out = self.final_activation(out)

        return out

    def model(self):
        x = tf.keras.layers.Input(shape=(self.height, self.width, 3))
        return tf.keras.Model(inputs=[x], outputs=self.call(x))