def linknet_inception_lite(input_shape):
inception_base = InceptionV3Same(input_shape=input_shape)
conv1 = inception_base.get_layer("activation_3").output
conv2 = inception_base.get_layer("activation_5").output
conv3 = inception_base.get_layer("mixed2").output
conv4 = inception_base.get_layer("mixed7").output
conv5 = inception_base.get_layer("mixed10").output
filters = [192, 288, 768, 2048]
feature_scale = 6
decoder5 = decoder(conv5,
filters[3],
filters[2],
name='decoder5',
feature_scale=feature_scale,
transposed_conv=True)
decoder5 = add([decoder5, conv4])
decoder4 = decoder(decoder5,
filters[2],
filters[1],
name='decoder4',
feature_scale=feature_scale,
transposed_conv=True)
decoder4 = add([decoder4, conv3])
decoder3 = decoder(decoder4,
filters[1],
filters[0],
name='decoder3',
feature_scale=feature_scale,
transposed_conv=True)
decoder3 = add([decoder3, conv2])
decoder2 = decoder(decoder3,
filters[0],
filters[0],
name='decoder2',
feature_scale=feature_scale,
transposed_conv=True)
decoder1 = concatenate([conv1, decoder2], axis=-1)
x = deconv_bn_relu(decoder1,
num_channels=32,
kernel_size=3,
name="decoder1",
transposed_conv=True)
x = linknet.conv_bn_relu(x, 32, 3, stride=1, padding='same', name='fc_1')
x = linknet.conv_bn_relu(x, 32, 3, stride=1, padding='same', name='fc_2')
x = SpatialDropout2D(0.5)(x)
x = Conv2D(1, (1, 1), activation="sigmoid")(x)
model = Model(inception_base.input, x)
return model
def linknet_resnet_50(input_shape):
resnet_base = resnet_50(input_shape=input_shape)
conv1 = resnet_base.get_layer("activation_1").output
conv2 = resnet_base.get_layer("activation_10").output
conv3 = resnet_base.get_layer("activation_22").output
conv4 = resnet_base.get_layer("activation_40").output
conv5 = resnet_base.get_layer("activation_49").output
filters = [256, 512, 1024, 2048]
feature_scale = 6
decoder5 = decoder(conv5,
filters[3],
filters[2],
name='decoder5',
feature_scale=feature_scale,
transposed_conv=True)
decoder5 = add([decoder5, conv4])
decoder4 = decoder(decoder5,
filters[2],
filters[1],
name='decoder4',
feature_scale=feature_scale,
transposed_conv=True)
decoder4 = add([decoder4, conv3])
decoder3 = decoder(decoder4,
filters[1],
filters[0],
name='decoder3',
feature_scale=feature_scale,
transposed_conv=True)
decoder3 = add([decoder3, conv2])
decoder2 = decoder(decoder3,
filters[0],
filters[0],
name='decoder2',
feature_scale=feature_scale,
transposed_conv=True)
decoder1 = concatenate([conv1, decoder2], axis=-1)
x = deconv_bn_relu(decoder1,
num_channels=32,
kernel_size=3,
name="decoder1",
transposed_conv=True)
x = linknet.conv_bn_relu(x, 32, 3, stride=1, padding='same', name='fc_1')
x = SpatialDropout2D(0.5)(x)
x = Conv2D(1, (1, 1), activation="sigmoid")(x)
model = Model(resnet_base.input, x)
return model
def linknet_inception_lite(input_shape):
inception_base = InceptionV3Same(input_shape=input_shape)
conv1 = inception_base.get_layer("activation_3").output
conv2 = inception_base.get_layer("activation_5").output
conv3 = inception_base.get_layer("mixed2").output
conv4 = inception_base.get_layer("mixed7").output
conv5 = inception_base.get_layer("mixed10").output
filters = [192, 288, 768, 2048]
feature_scale = 6
decoder5 = decoder(conv5, filters[3], filters[2], name='decoder5', feature_scale=feature_scale, transposed_conv=True)
decoder5 = add([decoder5, conv4])
decoder4 = decoder(decoder5, filters[2], filters[1], name='decoder4', feature_scale=feature_scale, transposed_conv=True)
decoder4 = add([decoder4, conv3])
decoder3 = decoder(decoder4, filters[1], filters[0], name='decoder3', feature_scale=feature_scale, transposed_conv=True)
decoder3 = add([decoder3, conv2])
decoder2 = decoder(decoder3, filters[0], filters[0], name='decoder2', feature_scale=feature_scale, transposed_conv=True)
decoder1 = concatenate([conv1, decoder2], axis=-1)
x = deconv_bn_relu(decoder1, num_channels=32, kernel_size=3, name="decoder1", transposed_conv=True)
x = linknet.conv_bn_relu(x, 32, 3, stride=1, padding='same', name='fc_1')
x = linknet.conv_bn_relu(x, 32, 3, stride=1, padding='same', name='fc_2')
x = SpatialDropout2D(0.5)(x)
x = Conv2D(1, (1, 1), activation="sigmoid")(x)
model = Model(inception_base.input, x)
return model
def linknet_resnet_50(input_shape):
resnet_base = resnet_50(input_shape=input_shape)
conv1 = resnet_base.get_layer("activation_1").output
conv2 = resnet_base.get_layer("activation_10").output
conv3 = resnet_base.get_layer("activation_22").output
conv4 = resnet_base.get_layer("activation_40").output
conv5 = resnet_base.get_layer("activation_49").output
filters = [256, 512, 1024, 2048]
feature_scale = 6
decoder5 = decoder(conv5, filters[3], filters[2], name='decoder5', feature_scale=feature_scale, transposed_conv=True)
decoder5 = add([decoder5, conv4])
decoder4 = decoder(decoder5, filters[2], filters[1], name='decoder4', feature_scale=feature_scale, transposed_conv=True)
decoder4 = add([decoder4, conv3])
decoder3 = decoder(decoder4, filters[1], filters[0], name='decoder3', feature_scale=feature_scale, transposed_conv=True)
decoder3 = add([decoder3, conv2])
decoder2 = decoder(decoder3, filters[0], filters[0], name='decoder2', feature_scale=feature_scale, transposed_conv=True)
decoder1 = concatenate([conv1, decoder2], axis=-1)
x = deconv_bn_relu(decoder1, num_channels=32, kernel_size=3, name="decoder1", transposed_conv=True)
x = linknet.conv_bn_relu(x, 32, 3, stride=1, padding='same', name='fc_1')
x = SpatialDropout2D(0.5)(x)
x = Conv2D(1, (1, 1), activation="sigmoid")(x)
model = Model(resnet_base.input, x)
return model
