def __init__(self, path, imsize, keras_mode=False):
self.imsize = imsize
self.keras_mode = keras_mode
self.images, self.ids = self.parse_data(path, keras_mode)
self.sample_ids = list(set([x[1:] for x in self.ids]))
self.sample_ids_train = self.sample_ids[:32]
self.sample_ids_val = self.sample_ids[32:64]
self.pre = vgg16.VGG16(include_top=False,
input_shape=(imsize, imsize, 3),
pooling='avg')
def VGG16(*args, **kwargs): return vgg16.VGG16(*args, **kwargs)
def get_gesunet(self):
# The first U-net:
#The fırst net is based on convolution with predefined filters. Number of filters is limited with sample size
inputs = Input((self.img_rows, self.img_cols, 1), name='first_data')
#data_format = 'channels_last'
#num_classes = 2
S = 9
K = (3, 3)
nf = 64
conv1 = Conv2D_LC(num_filters=nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(inputs)
conv1 = Conv2D_LC(num_filters=nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv1)
conv1 = Conv2D_LC(num_filters=nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D_LC(num_filters=2 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool1)
conv2 = Conv2D_LC(num_filters=2 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv2)
conv2 = Conv2D_LC(num_filters=2 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Conv2D_LC(num_filters=4 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool2)
conv3 = Conv2D_LC(num_filters=4 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv3)
conv3 = Conv2D_LC(num_filters=4 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Conv2D_LC(num_filters=8 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool3)
conv4 = Conv2D_LC(num_filters=8 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv4)
conv4 = Conv2D_LC(num_filters=8 * nf,
kernel_size=K,
sample_size=S,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv4)
up7 = Conv2D(4 * nf,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv4))
merge7 = concatenate([conv3, up7], axis=3)
conv7 = Conv2D(4 * nf,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge7)
conv7 = Conv2D(4 * nf,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv7)
up8 = Conv2D(2 * nf,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv7))
merge8 = concatenate([conv2, up8], axis=3)
conv8 = Conv2D(2 * nf,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge8)
conv8 = Conv2D(2 * nf,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv8)
up9 = Conv2D(nf,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv8))
merge9 = concatenate([conv1, up9], axis=3)
conv9 = Conv2D(nf,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge9)
conv9 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv9)
#conv9 = Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv9)
#conv9 = Conv2D(32, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv9)
#conv9 = Conv2D(16, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv9)
conv9 = Conv2D(1,
3,
activation='sigmoid',
padding='same',
kernel_initializer='he_normal')(conv9)
inputsT = concatenate(
[inputs, concatenate([inputs, conv9], axis=3)], axis=3)
modelA = Model(inputs=inputs, output=inputsT)
inputs2 = Input(shape=(128, 128, 3), name='Sec_data')
p2 = Input(shape=(4, 4, 1024))
vgg_conv = vgg16.VGG16(input_tensor=inputs2,
weights='imagenet',
include_top=False,
input_shape=(128, 128, 3),
classes=2)
# vgg_conv = vgg16.VGG16(input_tensor= imgs_train, weights='imagenet', include_top=False)
#imgs_train = vgg_conv.predict(imgs_train)
x1 = vgg_conv.get_layer('block1_conv2').output
x2 = vgg_conv.get_layer('block2_conv2').output
x3 = vgg_conv.get_layer('block3_conv3').output
x4 = vgg_conv.get_layer('block4_conv3').output
x5 = vgg_conv.get_layer('block5_pool').output
x6 = vgg_conv.get_layer('block5_conv3').output
conv5t = Conv2D(2048,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(x5)
conv5t = Conv2D(2048,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv5t)
up6t = Conv2D(1024,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv5t))
merge6t = concatenate([x6, up6t], axis=3)
conv6t = Conv2D(1024,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge6t)
conv6t = Conv2D(1024,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv6t)
up7t = Conv2D(512,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv6t))
merge7t = concatenate([x4, up7t], axis=3)
conv7t = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge7t)
conv7t = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv7t)
up8t = Conv2D(256,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv7t))
merge8t = concatenate([x3, up8t], axis=3)
conv8t = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge8t)
conv8t = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv8t)
up9t = Conv2D(128,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv8t))
merge9t = concatenate([x2, up9t], axis=3)
conv9t = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge9t)
conv9t = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv9t)
up10t = Conv2D(64,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv9t))
merge10t = concatenate([x1, up10t], axis=3)
conv10t = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge10t)
conv10t = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv10t)
conv10t = Conv2D(2,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv10t)
conv10t = Conv2D(1, 1, activation='sigmoid')(conv10t)
modelB = Model(inputs=inputT, output=conv10t)
Out1 = modelA(inputs)
Out = modelB(Out1)
modelC = Model(inputs=inputs, output=Out)
modelC.compile(optimizer=Adam(lr=0.000025),
loss="mean_squared_error",
metrics=['accuracy'])
return modelC