def _aspp(self, x, out_filters):
xs = list()
x1 = layers.Conv2D(out_filters, 1, strides=1)(x)
xs.append(x1)
for i in range(3):
xi = layers.Conv2D(out_filters,
3,
strides=1,
padding='same',
dilation_rate=6 * (i + 1))(x)
xs.append(xi)
img_pool = custom_layers.GlobalAveragePooling2D(keep_dims=True)(x)
img_pool = layers.Conv2D(out_filters,
1,
1,
kernel_initializer='he_normal')(img_pool)
img_pool = layers.UpSampling2D(size=self.aspp_size,
interpolation='bilinear')(img_pool)
xs.append(img_pool)
x = custom_layers.Concatenate(out_size=self.aspp_size)(xs)
x = layers.Conv2D(out_filters,
1,
strides=1,
kernel_initializer='he_normal')(x)
x = layers.BatchNormalization()(x)
return x
def _deeplab_v3_plus(self, inputs):
num_classes = self.num_classes
_, h, w, _ = backend.int_shape(inputs)
self.aspp_size = (h // 16, w // 16)
if self.base_model in [
'VGG16', 'VGG19', 'ResNet50', 'ResNet101', 'ResNet152',
'MobileNetV1', 'MobileNetV2'
]:
c2, c5 = self.encoder(inputs, output_stages=['c2', 'c5'])
else:
c2, c5 = self.encoder(inputs, output_stages=['c1', 'c5'])
x = self._aspp(c5, 256)
x = layers.Dropout(rate=0.5)(x)
x = layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(x)
x = self._conv_bn_relu(x, 48, 1, strides=1)
x = custom_layers.Concatenate(out_size=self.aspp_size)([x, c2])
x = self._conv_bn_relu(x, 256, 3, 1)
x = layers.Dropout(rate=0.5)(x)
x = self._conv_bn_relu(x, 256, 3, 1)
x = layers.Dropout(rate=0.1)(x)
x = layers.Conv2D(num_classes, 1, strides=1)(x)
x = layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(x)
outputs = x
return models.Model(inputs, outputs, name=self.version)
def _denseaspp(self, inputs):
_, inputs_h, inputs_w, _ = backend.int_shape(inputs)
aspp_size = inputs_h // 8, inputs_w // 8
num_classes = self.num_classes
c5 = self.encoder(inputs, output_stages='c5')
# First block rate=3
d3 = self._dilated_conv_block(c5, 256, 1)
d3 = self._dilated_conv_block(d3, 64, 3, rate=3)
# Second block rate=6
d4 = custom_layers.Concatenate(out_size=aspp_size)([c5, d3])
d4 = self._dilated_conv_block(d4, 256, 1)
d4 = self._dilated_conv_block(d4, 64, 3, rate=6)
# Third block rate=12
d5 = custom_layers.Concatenate(out_size=aspp_size)([c5, d3, d4])
d5 = self._dilated_conv_block(d5, 256, 1)
d5 = self._dilated_conv_block(d5, 64, 3, rate=12)
# Forth block rate=18
d6 = custom_layers.Concatenate(out_size=aspp_size)([c5, d3, d4, d5])
d6 = self._dilated_conv_block(d6, 256, 1)
d6 = self._dilated_conv_block(d6, 64, 3, rate=18)
# Fifth block rate=24
d7 = custom_layers.Concatenate(out_size=aspp_size)(
[c5, d3, d4, d5, d6])
d7 = self._dilated_conv_block(d7, 256, 1)
d7 = self._dilated_conv_block(d7, 64, 3, rate=24)
x = custom_layers.Concatenate(out_size=aspp_size)(
[c5, d3, d4, d5, d6, d7])
x = layers.Conv2D(num_classes,
1,
strides=1,
kernel_initializer='he_normal')(x)
x = layers.UpSampling2D(size=(8, 8), interpolation='bilinear')(x)
outputs = x
return models.Model(inputs, outputs, name=self.version)