def _aspp(self, x, out_filters):
xs = list()
x1 = layers.Conv2D(out_filters,
1,
strides=1,
kernel_initializer='he_normal')(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 _attention_refinement_module(self, x):
# Global average pooling
_, _, _, c = backend.int_shape(x)
glb = custom_layers.GlobalAveragePooling2D(keep_dims=True)(x)
glb = layers.Conv2D(c, 1, strides=1, kernel_initializer='he_normal')(glb)
glb = layers.BatchNormalization()(glb)
glb = layers.Activation(activation='sigmoid')(glb)
x = layers.Multiply()([x, glb])
return x
def _feature_fusion_module(self, input_1, input_2, filters):
inputs = layers.Concatenate()([input_1, input_2])
inputs = self._conv_block(inputs, filters=filters, kernel_size=3)
# Global average pooling
_, _, _, c = backend.int_shape(inputs)
glb = custom_layers.GlobalAveragePooling2D(keep_dims=True)(inputs)
glb = layers.Conv2D(filters, 1, strides=1, activation='relu', kernel_initializer='he_normal')(glb)
glb = layers.Conv2D(filters, 1, strides=1, activation='sigmoid', kernel_initializer='he_normal')(glb)
x = layers.Multiply()([inputs, glb])
return x
def _gau(self, x, y, out_filters, up_size=(2, 2)):
glb = custom_layers.GlobalAveragePooling2D(keep_dims=True)(y)
glb = layers.Conv2D(out_filters,
1,
strides=1,
activation='sigmoid',
kernel_initializer='he_normal')(glb)
x = self._conv_bn_relu(x, out_filters, 3, 1)
x = layers.Multiply()([x, glb])
y = layers.UpSampling2D(size=up_size, interpolation='bilinear')(y)
y = layers.Add()([x, y])
return y
def _fpa(self, x, out_filters):
_, h, w, _ = backend.int_shape(x)
# global average pooling
glb = custom_layers.GlobalAveragePooling2D(keep_dims=True)(x)
glb = layers.Conv2D(out_filters,
1,
strides=1,
kernel_initializer='he_normal')(glb)
# down
down1 = layers.AveragePooling2D(pool_size=(2, 2))(x)
down1 = self._conv_bn_relu(down1, 1, 7, 1)
down2 = layers.AveragePooling2D(pool_size=(2, 2))(down1)
down2 = self._conv_bn_relu(down2, 1, 5, 1)
down3 = layers.AveragePooling2D(pool_size=(2, 2))(down2)
down3 = self._conv_bn_relu(down3, 1, 3, 1)
down1 = self._conv_bn_relu(down1, 1, 7, 1)
down2 = self._conv_bn_relu(down2, 1, 5, 1)
down3 = self._conv_bn_relu(down3, 1, 3, 1)
# up
up2 = layers.UpSampling2D(size=(2, 2))(down3)
up2 = layers.Add()([up2, down2])
up1 = layers.UpSampling2D(size=(2, 2))(up2)
up1 = layers.Add()([up1, down1])
up = layers.UpSampling2D(size=(2, 2))(up1)
x = layers.Conv2D(out_filters,
1,
strides=1,
kernel_initializer='he_normal')(x)
x = layers.BatchNormalization()(x)
# multiply
x = layers.Multiply()([x, up])
# add
x = layers.Add()([x, glb])
return x
def _bisegnet(self, inputs):
num_classes = self.num_classes
# the spatial path
sx = self._conv_block(inputs, 64, 3, 2)
sx = self._conv_block(sx, 128, 3, 2)
sx = self._conv_block(sx, 256, 3, 2)
# the context path
if self.base_model in ['VGG16',
'VGG19',
'ResNet50',
'ResNet101',
'ResNet152',
'MobileNetV1',
'MobileNetV2',
'Xception',
'Xception-DeepLab']:
c4, c5 = self.encoder(inputs, output_stages=['c4', 'c5'])
else:
c4, c5 = self.encoder(inputs, output_stages=['c3', 'c5'])
c4 = self._attention_refinement_module(c4)
c5 = self._attention_refinement_module(c5)
glb = custom_layers.GlobalAveragePooling2D(keep_dims=True)(c5)
c5 = layers.Multiply()([c5, glb])
# combining the paths
c4 = layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(c4)
c5 = layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(c5)
cx = layers.Concatenate()([c4, c5])
x = self._feature_fusion_module(sx, cx, num_classes)
x = layers.UpSampling2D(size=(8, 8), interpolation='bilinear')(x)
x = layers.Conv2D(num_classes, 1, 1, kernel_initializer='he_normal')(x)
outputs = x
return models.Model(inputs, outputs, name=self.version)
def _pspnet(self, inputs):
num_classes = self.num_classes
_, inputs_h, inputs_w, _ = backend.int_shape(inputs)
h, w = inputs_h // 8, inputs_w // 8
x = self.encoder(inputs)
if not (h % 6 == 0 and w % 6 == 0):
raise ValueError(
'\'pyramid pooling\' size must be divided by 6, but received {size}'
.format(size=(h, w)))
pool_size = [(h, w), (h // 2, w // 2), (h // 3, w // 3),
(h // 6, w // 6)]
# pyramid pooling
x1 = custom_layers.GlobalAveragePooling2D(keep_dims=True)(x)
x1 = layers.Conv2D(512, 1, strides=1,
kernel_initializer='he_normal')(x1)
x1 = layers.BatchNormalization()(x1)
x1 = layers.ReLU()(x1)
x1 = layers.UpSampling2D(size=pool_size[0])(x1)
x2 = layers.AveragePooling2D(pool_size=pool_size[1])(x)
x2 = layers.Conv2D(512, 1, strides=1,
kernel_initializer='he_normal')(x2)
x2 = layers.BatchNormalization()(x2)
x2 = layers.ReLU()(x2)
x2 = layers.UpSampling2D(size=pool_size[1])(x2)
x3 = layers.AveragePooling2D(pool_size=pool_size[2])(x)
x3 = layers.Conv2D(512, 1, strides=1,
kernel_initializer='he_normal')(x3)
x3 = layers.BatchNormalization()(x3)
x3 = layers.ReLU()(x3)
x3 = layers.UpSampling2D(size=pool_size[2])(x3)
x6 = layers.AveragePooling2D(pool_size=pool_size[3])(x)
x6 = layers.Conv2D(512, 1, strides=1,
kernel_initializer='he_normal')(x6)
x6 = layers.BatchNormalization()(x6)
x6 = layers.ReLU()(x6)
x6 = layers.UpSampling2D(size=pool_size[3])(x6)
x = layers.Concatenate()([x, x1, x2, x3, x6])
x = layers.Conv2D(512,
3,
strides=1,
padding='same',
kernel_initializer='he_normal')(x)
x = layers.BatchNormalization()(x)
x = layers.ReLU()(x)
x = layers.Conv2D(num_classes,
1,
strides=1,
kernel_initializer='he_normal')(x)
x = layers.BatchNormalization()(x)
x = layers.UpSampling2D(size=(8, 8), interpolation='bilinear')(x)
outputs = x
return models.Model(inputs, outputs, name=self.version)