def build_model(self):
input = Input(batch_shape=self.input_shape, name='input_1')
conv1_1 = Convolution2D(64, 3, 3, activation='relu')(input)
conv1_2 = Convolution2D(64, 3, 3, activation='relu')(conv1_1)
conv1_out = MaxPooling2D()(conv1_2)
conv2_1 = Convolution2D(128, 3, 3, activation='relu')(conv1_out)
dropout2 = Dropout(0.5)(conv2_1)
conv2_2 = Convolution2D(128, 3, 3, activation='relu')(dropout2)
conv2_out = MaxPooling2D()(conv2_2)
conv3_1 = Convolution2D(256, 3, 3, activation='relu')(conv2_out)
dropout3 = Dropout(0.5)(conv3_1)
conv3_2 = Convolution2D(256, 3, 3, activation='relu')(dropout3)
conv3_out = MaxPooling2D()(conv3_2)
conv4_1 = Convolution2D(512, 3, 3, activation='relu')(conv3_out)
dropout4 = Dropout(0.5)(conv4_1)
conv4_2 = Convolution2D(512, 3, 3, activation='relu')(dropout4)
conv4_out = MaxPooling2D()(conv4_2)
conv5_1 = Convolution2D(1024, 3, 3, activation='relu')(conv4_out)
dropout5 = Dropout(0.5)(conv5_1)
conv5_2 = Convolution2D(1024, 3, 3, activation='relu')(dropout5)
conv5_out = MaxPooling2D()(conv5_2)
up_conv1 = self.upconv2_2(conv5_out, conv4_out, 512)
# conv6_out = self.conv3_3(up_conv1, 512)
up_conv2 = self.upconv2_2(up_conv1, conv3_out, 256)
# conv7_out = self.conv3_3(up_conv2, 256)
up_conv3 = self.upconv2_2(up_conv2, conv2_out, 128)
# conv8_out = self.conv3_3(up_conv3, 128)
up_conv4 = self.upconv2_2(up_conv3, conv1_out, 64)
# conv9_out = self.conv3_3(up_conv4, 64)
out_shape = [dim.value for dim in input.get_shape()]
out_shape = [self.batch_size] + out_shape[1:3] + [self.no_classes]
output = Deconvolution2D(self.no_classes,
5,
5,
out_shape,
subsample=(2, 2),
activation='sigmoid',
name='class_out')(up_conv4)
model = Model(input, output)
optimizer = Adam(lr=self.learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
model.compile(optimizer=optimizer,
loss={'class_out': 'binary_crossentropy'},
metrics=['binary_accuracy'])
if self.weight_file:
model.load_weights(self.weight_file)
model.summary()
return model
def build_model(self):
input = Input(batch_shape=self.input_shape, name='input_1')
fileter_size = 3
# Block 1
conv1_1 = Convolution2D(64,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv1_1')(input)
conv1_2 = Convolution2D(64,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv1_2')(conv1_1)
conv1_out = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
border_mode='same',
name='pool1')(conv1_2)
# Block 2
conv2_1 = Convolution2D(128,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv2_1')(conv1_out)
conv2_2 = Convolution2D(128,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv2_2')(conv2_1)
conv2_out = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
border_mode='same',
name='pool2')(conv2_2)
# Block 3
conv3_1 = Convolution2D(256,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv3_1')(conv2_out)
conv3_2 = Convolution2D(256,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv3_2')(conv3_1)
conv3_3 = Convolution2D(256,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv3_3')(conv3_2)
conv3_out = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
border_mode='same',
name='pool3')(conv3_3)
# Block 4
conv4_1 = Convolution2D(512,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv4_1')(conv3_out)
conv4_2 = Convolution2D(512,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv4_2')(conv4_1)
conv4_3 = Convolution2D(512,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv4_3')(conv4_2)
conv4_out = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
border_mode='same',
name='pool4')(conv4_3)
# Block 5
conv5_1 = Convolution2D(512,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv5_1')(conv4_out)
conv5_2 = Convolution2D(512,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv5_2')(conv5_1)
conv5_3 = Convolution2D(512,
fileter_size,
fileter_size,
activation='relu',
border_mode='same',
name='conv5_3')(conv5_2)
conv5_out = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
border_mode='same',
name='pool5')(conv5_3)
# Block 6
conv6_1 = Convolution2D(4096,
7,
7,
activation='relu',
border_mode='same',
name='conv6_1')(conv5_out)
conv6_out = Dropout(0.5)(conv6_1)
# Block 7
conv7_1 = Convolution2D(4096,
1,
1,
activation='relu',
border_mode='same',
name='conv7_1')(conv6_out)
conv7_out = Dropout(0.5)(conv7_1)
# De1
score_conv7_out = Convolution2D(self.no_classes,
1,
1,
border_mode='same')(conv7_out)
score_pool4 = Convolution2D(self.no_classes, 1, 1,
border_mode='same')(conv4_out)
out_shape = [dim.value for dim in score_pool4.get_shape()]
up_conv_1 = Deconvolution2D(self.no_classes,
4,
4,
output_shape=out_shape,
border_mode="same",
subsample=(2, 2))(score_conv7_out)
upscore_1 = merge([score_pool4, up_conv_1], mode='sum', concat_axis=-1)
# De2
score_pool3 = Convolution2D(self.no_classes, 1, 1,
border_mode='same')(conv3_out)
out_shape = [dim.value for dim in score_pool3.get_shape()]
up_conv_2 = Deconvolution2D(self.no_classes,
4,
4,
output_shape=out_shape,
border_mode="same",
subsample=(2, 2))(upscore_1)
upscore_2 = merge([score_pool3, up_conv_2], mode='sum', concat_axis=-1)
# up_conv1 = self.upconv2_2(conv7_out, conv4_out, 512)
# conv6_out = self.convfileter_size_fileter_size(up_conv1, 512)
# up_conv2 = self.upconv2_2(up_conv1, conv3_out, 256)
# conv7_out = self.convfileter_size_fileter_size(up_conv2, 256)
out_shape = [dim.value for dim in input.get_shape()]
out_shape = [self.batch_size
] + out_shape[1:fileter_size] + [self.no_classes]
output = Deconvolution2D(self.no_classes,
16,
16,
output_shape=out_shape,
border_mode="same",
subsample=(8, 8))(upscore_2)
output = Reshape((self.input_shape[1] * self.input_shape[2],
self.no_classes))(output)
output = Activation(activation='softmax', name='class_out')(output)
model = Model(input, output)
return model