def preforward():
# input
inputs1 = l = Input((9, 9, 3), name='inputs1')
# conv1
l_offset = ConvOffset2D(3, name='conv1_offset')(l)
l = Conv2D(128, (3, 3), padding='same', strides=(2, 2), name='conv1', trainable=train)(l_offset)
l = Activation('relu', name='conv1_relu')(l)
l = MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None, name='max_pool_1')(l)
# conv2
l_offset = ConvOffset2D(128, name='conv2_offset')(l)
l = Conv2D(256, (3, 3), padding='same', name='conv2', trainable=train)(l_offset)
l = Activation('relu', name='conv2_relu')(l)
l = MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None, name='max_pool_2')(l)
l = Flatten()(l)
l = Dense(256, name='fc1', trainable=train)(l)
# ***************************************************
# output
# ***************************************************
l = Dense(256, activation='sigmoid')(l)
outputs = Dense(2, activation='softmax')(l) #输出类型要改！！！
model = Model(inputs=inputs1, outputs=outputs)
optim = Adam()
# optim = SGD(1e-3, momentum=0.99, nesterov=True)
loss = categorical_crossentropy
model.compile(optim, loss, metrics=['accuracy'])
return model
def __init__(self, ninput, dilated, k=40, dropprob=0.02):
super(EDABlock, self).__init__()
#k: growthrate
#dropprob:a dropout layer between the last ReLU and the concatenation of each module
self.conv1x1 = nn.Conv2d(ninput, k, kernel_size=1)
self.bn0 = nn.BatchNorm2d(k)
self.offset1 = ConvOffset2D(filters=k) #kernel_size=(1, 1)
self.conv3x1_1 = nn.Conv2d(k, k, kernel_size=(3, 1), padding=(1, 0))
self.conv1x3_1 = nn.Conv2d(k, k, kernel_size=(1, 3), padding=(0, 1))
self.bn1 = nn.BatchNorm2d(k)
self.offset2 = ConvOffset2D(
filters=k
) #kernel_size=(3, 1),stride=1, padding=(dilated,0), dilation = dilated
self.conv3x1_2 = nn.Conv2d(k,
k, (3, 1),
stride=1,
padding=(dilated, 0),
dilation=dilated)
self.conv1x3_2 = nn.Conv2d(k,
k, (1, 3),
stride=1,
padding=(0, dilated),
dilation=dilated)
self.bn2 = nn.BatchNorm2d(k)
self.dropout = nn.Dropout2d(dropprob)
def get_deform_cnn(trainable):
inputs = l = Input((28, 28, 1), name='input')
# conv11
l = Conv2D(32, (3, 3), padding='same', name='conv11',
trainable=trainable)(l)
l = Activation('relu', name='conv11_relu')(l)
l = BatchNormalization(name='conv11_bn')(l)
# conv12
l_offset = ConvOffset2D(32, name='conv12_offset')(l)
l = Conv2D(64, (3, 3),
padding='same',
strides=(2, 2),
name='conv12',
trainable=trainable)(l_offset)
l = Activation('relu', name='conv12_relu')(l)
l = BatchNormalization(name='conv12_bn')(l)
# conv21
l_offset = ConvOffset2D(64, name='conv21_offset')(l)
l = Conv2D(128, (3, 3), padding='same', name='conv21',
trainable=trainable)(l_offset)
l = Activation('relu', name='conv21_relu')(l)
l = BatchNormalization(name='conv21_bn')(l)
# conv22
l_offset = ConvOffset2D(128, name='conv22_offset')(l)
l = Conv2D(128, (3, 3),
padding='same',
strides=(2, 2),
name='conv22',
trainable=trainable)(l_offset)
l = Activation('relu', name='conv22_relu')(l)
l = BatchNormalization(name='conv22_bn')(l)
# out
l = GlobalAvgPool2D(name='avg_pool')(l)
l = Dense(10, name='fc1', trainable=trainable)(l)
outputs = l = Activation('softmax', name='out')(l)
return inputs, outputs
def __init__(self):
super(DeformConvNet, self).__init__()
# conv11
self.conv11 = nn.Conv2d(3, 32, 3, padding=1)
self.bn11 = nn.BatchNorm2d(32)
# conv12
self.offset12 = ConvOffset2D(32)
self.conv12 = nn.Conv2d(32, 64, 3, padding=1, stride=2)
self.bn12 = nn.BatchNorm2d(64)
# conv21
self.offset21 = ConvOffset2D(64)
self.conv21 = nn.Conv2d(64, 128, 3, padding=1)
self.bn21 = nn.BatchNorm2d(128)
# conv22
self.offset22 = ConvOffset2D(128)
self.conv22 = nn.Conv2d(128, 128, 3, padding=1, stride=2)
self.bn22 = nn.BatchNorm2d(128)
# out
self.fc = nn.Linear(128, 128)
def InceptionV3_deform_model(nb_classes):
if K.image_data_format() == 'channels_first':
channel_axis = 1
else:
channel_axis = 3
img_input = InceptionV3().input
x = conv2d_bn(img_input, 32, 3, 3, strides=(2, 2), padding='valid')
x = ConvOffset2D(32, name='conv1_offset')(x)
x = conv2d_bn(x, 32, 3, 3, padding='valid')
x = ConvOffset2D(32, name='conv2_offset')(x)
x = conv2d_bn(x, 64, 3, 3)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = conv2d_bn(x, 80, 1, 1, padding='valid')
x = ConvOffset2D(80, name='conv3_offset')(x)
x = conv2d_bn(x, 192, 3, 3, padding='valid')
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
# mixed 0, 1, 2: 35 x 35 x 256
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv4_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv5_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 32, 1, 1)
x = concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed0')
# mixed 1: 35 x 35 x 256
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv6_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv7_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 64, 1, 1)
x = concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed1')
# mixed 2: 35 x 35 x 256
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv8_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv9_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 64, 1, 1)
x = concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed2')
# mixed 3: 17 x 17 x 768
x = ConvOffset2D(288, name='conv10_offset')(x)
branch3x3 = conv2d_bn(x, 384, 3, 3, strides=(2, 2), padding='valid')
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv11_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv12_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl,
96,
3,
3,
strides=(2, 2),
padding='valid')
branch_pool = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = concatenate([branch3x3, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed3')
# mixed 4: 17 x 17 x 768
branch1x1 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(x, 128, 1, 1)
branch7x7 = conv2d_bn(branch7x7, 128, 1, 7)
branch7x7 = conv2d_bn(branch7x7, 192, 7, 1)
branch7x7dbl = conv2d_bn(x, 128, 1, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 1, 7)
branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=channel_axis,
name='mixed4')
# mixed 5, 6: 17 x 17 x 768
for i in range(2):
branch1x1 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(x, 160, 1, 1)
branch7x7 = conv2d_bn(branch7x7, 160, 1, 7)
branch7x7 = conv2d_bn(branch7x7, 192, 7, 1)
branch7x7dbl = conv2d_bn(x, 160, 1, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 1, 7)
branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=channel_axis,
name='mixed' + str(5 + i))
# mixed 7: 17 x 17 x 768
branch1x1 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(branch7x7, 192, 1, 7)
branch7x7 = conv2d_bn(branch7x7, 192, 7, 1)
branch7x7dbl = conv2d_bn(x, 192, 1, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=channel_axis,
name='mixed7')
# mixed 8: 8 x 8 x 1280
branch3x3 = conv2d_bn(x, 192, 1, 1)
branch3x3 = ConvOffset2D(192, name='conv13_offset')(branch3x3)
branch3x3 = conv2d_bn(branch3x3,
320,
3,
3,
strides=(2, 2),
padding='valid')
branch7x7x3 = conv2d_bn(x, 192, 1, 1)
branch7x7x3 = conv2d_bn(branch7x7x3, 192, 1, 7)
branch7x7x3 = conv2d_bn(branch7x7x3, 192, 7, 1)
branch7x7x3 = ConvOffset2D(192, name='conv14_offset')(branch7x7x3)
branch7x7x3 = conv2d_bn(branch7x7x3,
192,
3,
3,
strides=(2, 2),
padding='valid')
branch_pool = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = concatenate([branch3x3, branch7x7x3, branch_pool],
axis=channel_axis,
name='mixed8')
# mixed 9: 8 x 8 x 2048
for i in range(2):
branch1x1 = conv2d_bn(x, 320, 1, 1)
branch3x3 = conv2d_bn(x, 384, 1, 1)
branch3x3_1 = conv2d_bn(branch3x3, 384, 1, 3)
branch3x3_2 = conv2d_bn(branch3x3, 384, 3, 1)
branch3x3 = concatenate([branch3x3_1, branch3x3_2],
axis=channel_axis,
name='mixed9_' + str(i))
branch3x3dbl = conv2d_bn(x, 448, 1, 1)
branch3x3dbl = ConvOffset2D(448, name='conv15_%d_offset' %
(i))(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 384, 3, 3)
branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 384, 1, 3)
branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 384, 3, 1)
branch3x3dbl = concatenate([branch3x3dbl_1, branch3x3dbl_2],
axis=channel_axis)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch3x3, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed' + str(9 + i))
# Classification block
x = GlobalAveragePooling2D(name='avg_pool')(x)
x = Dense(nb_classes, activation='softmax', name='predictions')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
inputs = img_input
# Create model.
model = Model(inputs, x, name='inception_v3_deform')
return model
def deform_cnn(input_tensor=None,
input_shape=None,
classes=10,
trainable=True):
if K.image_data_format() == 'channels_first':
input_shape = (1, img_rows, img_cols)
bn_axis = 1
else:
input_shape = (img_rows, img_cols, 1)
bn_axis = 3
if input_tensor is None:
inputs = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
inputs = Input(tensor=input_tensor, shape=input_shape)
else:
inputs = input_tensor
#Conv_1 layer
x = Conv2D(32, (3, 3), padding='same', name='conv_1',
trainable=trainable)(inputs)
x = BatchNormalization(axis=bn_axis)(x)
x = Activation('relu')(x)
#Conv_2 layer
x_offset = ConvOffset2D(32, name='conv_2_offset')(x)
x = Conv2D(64, (3, 3),
strides=(2, 2),
padding='same',
name='conv_2',
trainable=trainable)(x_offset)
x = BatchNormalization(axis=bn_axis)(x)
x = Activation('relu')(x)
#Conv_3 layer
x_offset = ConvOffset2D(64, name='conv_3_offset')(x)
x = Conv2D(128, (3, 3),
strides=(2, 2),
padding='same',
name='conv_3',
trainable=trainable)(x_offset)
x = BatchNormalization(axis=bn_axis)(x)
x = Activation('relu')(x)
#Conv_4 layer
x_offset = ConvOffset2D(128, name='conv_4_offset')(x)
x = Conv2D(128, (3, 3), padding='same', name='conv_4',
trainable=trainable)(x_offset)
x = BatchNormalization(axis=bn_axis)(x)
x = Activation('relu')(x)
#Pooling layer
x = GlobalAveragePooling2D()(x)
#fc layer
outputs = Dense(classes,
activation='softmax',
name='fc',
trainable=trainable)(x)
return inputs, outputs
def build_deformable_bn(width, height, n_classes, weights_path=None, train=False):
inp = Input(shape=(height, width, 3))
#regularized
x = Lambda(lambda x: (x - 127.5)/255.0)(inp)
# (1/2)
y = Lambda(lambda x: tf.image.resize_bilinear(x, size=(int(x.shape[1])//2, int(x.shape[2])//2)), name='data_sub2')(x)
y = Conv2D(32, 3, strides=2, padding='same', activation='relu', name='conv1_1_3x3_s2')(y)
y = BatchNormalization(name='conv1_1_3x3_s2_bn')(y)
y = ConvOffset2D(32, name="conv1_1_offset")(y)
y = Conv2D(32, 3, padding='same', activation='relu', name='conv1_2_3x3')(y)
y = BatchNormalization(name='conv1_2_3x3_s2_bn')(y)
y = ConvOffset2D(32, name="conv1_2_offset")(y)
y = Conv2D(64, 3, padding='same', activation='relu', name='conv1_3_3x3')(y)
y = BatchNormalization(name='conv1_3_3x3_bn')(y)
y_ = MaxPooling2D(pool_size=3, strides=2, name='pool1_3x3_s2')(y)
y = Conv2D(128, 1, name='conv2_1_1x1_proj')(y_)
y = BatchNormalization(name='conv2_1_1x1_proj_bn')(y)
y_ = Conv2D(32, 1, activation='relu', name='conv2_1_1x1_reduce')(y_)
y_ = BatchNormalization(name='conv2_1_1x1_reduce_bn')(y_)
y_ = ZeroPadding2D(name='padding1')(y_)
y_ = Conv2D(32, 3, activation='relu', name='conv2_1_3x3')(y_)
y_ = BatchNormalization(name='conv2_1_3x3_bn')(y_)
y_ = Conv2D(128, 1, name='conv2_1_1x1_increase')(y_)
y_ = BatchNormalization(name='conv2_1_1x1_increase_bn')(y_)
y = Add(name='conv2_1')([y,y_])
y_ = Activation('relu', name='conv2_1/relu')(y)
y = Conv2D(32, 1, activation='relu', name='conv2_2_1x1_reduce')(y_)
y = BatchNormalization(name='conv2_2_1x1_reduce_bn')(y)
y = ZeroPadding2D(name='padding2')(y)
y = Conv2D(32, 3, activation='relu', name='conv2_2_3x3')(y)
y = BatchNormalization(name='conv2_2_3x3_bn')(y)
y = Conv2D(128, 1, name='conv2_2_1x1_increase')(y)
y = BatchNormalization(name='conv2_2_1x1_increase_bn')(y)
y = Add(name='conv2_2')([y,y_])
y_ = Activation('relu', name='conv2_2/relu')(y)
y = Conv2D(32, 1, activation='relu', name='conv2_3_1x1_reduce')(y_)
y = BatchNormalization(name='conv2_3_1x1_reduce_bn')(y)
y = ZeroPadding2D(name='padding3')(y)
y = Conv2D(32, 3, activation='relu', name='conv2_3_3x3')(y)
y = BatchNormalization(name='conv2_3_3x3_bn')(y)
y = Conv2D(128, 1, name='conv2_3_1x1_increase')(y)
y = BatchNormalization(name='conv2_3_1x1_increase_bn')(y)
y = Add(name='conv2_3')([y,y_])
y_ = Activation('relu', name='conv2_3/relu')(y)
y = Conv2D(256, 1, strides=2, name='conv3_1_1x1_proj')(y_)
y = BatchNormalization(name='conv3_1_1x1_proj_bn')(y)
y_ = Conv2D(64, 1, strides=2, activation='relu', name='conv3_1_1x1_reduce')(y_)
y_ = BatchNormalization(name='conv3_1_1x1_reduce_bn')(y_)
y_ = ZeroPadding2D(name='padding4')(y_)
y_ = Conv2D(64, 3, activation='relu', name='conv3_1_3x3')(y_)
y_ = BatchNormalization(name='conv3_1_3x3_bn')(y_)
y_ = Conv2D(256, 1, name='conv3_1_1x1_increase')(y_)
y_ = BatchNormalization(name='conv3_1_1x1_increase_bn')(y_)
y = Add(name='conv3_1')([y,y_])
z = Activation('relu', name='conv3_1/relu')(y)
# (1/4)
y_ = Lambda(lambda x: tf.image.resize_bilinear(x, size=(int(x.shape[1])//2, int(x.shape[2])//2)), name='conv3_1_sub4')(z)
y = Conv2D(64, 1, activation='relu', name='conv3_2_1x1_reduce')(y_)
y = BatchNormalization(name='conv3_2_1x1_reduce_bn')(y)
y = ZeroPadding2D(name='padding5')(y)
y = ConvOffset2D(64, name="conv3_2_1_Offset")(y)
y = Conv2D(64, 3, activation='relu', name='conv3_2_3x3')(y)
y = BatchNormalization(name='conv3_2_3x3_bn')(y)
y = Conv2D(256, 1, name='conv3_2_1x1_increase')(y)
y = BatchNormalization(name='conv3_2_1x1_increase_bn')(y)
y = Add(name='conv3_2')([y,y_])
y_ = Activation('relu', name='conv3_2/relu')(y)
y = Conv2D(64, 1, activation='relu', name='conv3_3_1x1_reduce')(y_)
y = BatchNormalization(name='conv3_3_1x1_reduce_bn')(y)
y = ZeroPadding2D(name='padding6')(y)
y = Conv2D(64, 3, activation='relu', name='conv3_3_3x3')(y)
y = BatchNormalization(name='conv3_3_3x3_bn')(y)
y = Conv2D(256, 1, name='conv3_3_1x1_increase')(y)
y = BatchNormalization(name='conv3_3_1x1_increase_bn')(y)
y = Add(name='conv3_3')([y,y_])
y_ = Activation('relu', name='conv3_3/relu')(y)
y = Conv2D(64, 1, activation='relu', name='conv3_4_1x1_reduce')(y_)
y = BatchNormalization(name='conv3_4_1x1_reduce_bn')(y)
y = ZeroPadding2D(name='padding7')(y)
y = Conv2D(64, 3, activation='relu', name='conv3_4_3x3')(y)
y = BatchNormalization(name='conv3_4_3x3_bn')(y)
y = Conv2D(256, 1, name='conv3_4_1x1_increase')(y)
y = BatchNormalization(name='conv3_4_1x1_increase_bn')(y)
y = Add(name='conv3_4')([y,y_])
y_ = Activation('relu', name='conv3_4/relu')(y)
y = Conv2D(512, 1, name='conv4_1_1x1_proj')(y_)
y = BatchNormalization(name='conv4_1_1x1_proj_bn')(y)
y_ = Conv2D(128, 1, activation='relu', name='conv4_1_1x1_reduce')(y_)
y_ = BatchNormalization(name='conv4_1_1x1_reduce_bn')(y_)
y_ = ZeroPadding2D(padding=2, name='padding8')(y_)
y_ = Conv2D(128, 3, dilation_rate=2, activation='relu', name='conv4_1_3x3')(y_)
y_ = BatchNormalization(name='conv4_1_3x3_bn')(y_)
y_ = Conv2D(512, 1, name='conv4_1_1x1_increase')(y_)
y_ = BatchNormalization(name='conv4_1_1x1_increase_bn')(y_)
y = Add(name='conv4_1')([y,y_])
y_ = Activation('relu', name='conv4_1/relu')(y)
y = Conv2D(128, 1, activation='relu', name='conv4_2_1x1_reduce')(y_)
y = BatchNormalization(name='conv4_2_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=2, name='padding9')(y)
y = Conv2D(128, 3, dilation_rate=2, activation='relu', name='conv4_2_3x3')(y)
y = BatchNormalization(name='conv4_2_3x3_bn')(y)
y = Conv2D(512, 1, name='conv4_2_1x1_increase')(y)
y = BatchNormalization(name='conv4_2_1x1_increase_bn')(y)
y = Add(name='conv4_2')([y,y_])
y_ = Activation('relu', name='conv4_2/relu')(y)
y = Conv2D(128, 1, activation='relu', name='conv4_3_1x1_reduce')(y_)
y = BatchNormalization(name='conv4_3_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=2, name='padding10')(y)
y = Conv2D(128, 3, dilation_rate=2, activation='relu', name='conv4_3_3x3')(y)
y = BatchNormalization(name='conv4_3_3x3_bn')(y)
y = Conv2D(512, 1, name='conv4_3_1x1_increase')(y)
y = BatchNormalization(name='conv4_3_1x1_increase_bn')(y)
y = Add(name='conv4_3')([y,y_])
y_ = Activation('relu', name='conv4_3/relu')(y)
y = Conv2D(128, 1, activation='relu', name='conv4_4_1x1_reduce')(y_)
y = BatchNormalization(name='conv4_4_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=2, name='padding11')(y)
y = Conv2D(128, 3, dilation_rate=2, activation='relu', name='conv4_4_3x3')(y)
y = BatchNormalization(name='conv4_4_3x3_bn')(y)
y = Conv2D(512, 1, name='conv4_4_1x1_increase')(y)
y = BatchNormalization(name='conv4_4_1x1_increase_bn')(y)
y = Add(name='conv4_4')([y,y_])
y_ = Activation('relu', name='conv4_4/relu')(y)
y = Conv2D(128, 1, activation='relu', name='conv4_5_1x1_reduce')(y_)
y = BatchNormalization(name='conv4_5_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=2, name='padding12')(y)
y = Conv2D(128, 3, dilation_rate=2, activation='relu', name='conv4_5_3x3')(y)
y = BatchNormalization(name='conv4_5_3x3_bn')(y)
y = Conv2D(512, 1, name='conv4_5_1x1_increase')(y)
y = BatchNormalization(name='conv4_5_1x1_increase_bn')(y)
y = Add(name='conv4_5')([y,y_])
y_ = Activation('relu', name='conv4_5/relu')(y)
y = Conv2D(128, 1, activation='relu', name='conv4_6_1x1_reduce')(y_)
y = BatchNormalization(name='conv4_6_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=2, name='padding13')(y)
y = Conv2D(128, 3, dilation_rate=2, activation='relu', name='conv4_6_3x3')(y)
y = BatchNormalization(name='conv4_6_3x3_bn')(y)
y = Conv2D(512, 1, name='conv4_6_1x1_increase')(y)
y = BatchNormalization(name='conv4_6_1x1_increase_bn')(y)
y = Add(name='conv4_6')([y,y_])
y = Activation('relu', name='conv4_6/relu')(y)
y_ = Conv2D(1024, 1, name='conv5_1_1x1_proj')(y)
y_ = BatchNormalization(name='conv5_1_1x1_proj_bn')(y_)
y = Conv2D(256, 1, activation='relu', name='conv5_1_1x1_reduce')(y)
y = BatchNormalization(name='conv5_1_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=4, name='padding14')(y)
y = Conv2D(256, 3, dilation_rate=4, activation='relu', name='conv5_1_3x3')(y)
y = BatchNormalization(name='conv5_1_3x3_bn')(y)
y = Conv2D(1024, 1, name='conv5_1_1x1_increase')(y)
y = BatchNormalization(name='conv5_1_1x1_increase_bn')(y)
y = Add(name='conv5_1')([y,y_])
y_ = Activation('relu', name='conv5_1/relu')(y)
y = Conv2D(256, 1, activation='relu', name='conv5_2_1x1_reduce')(y_)
y = BatchNormalization(name='conv5_2_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=4, name='padding15')(y)
y = Conv2D(256, 3, dilation_rate=4, activation='relu', name='conv5_2_3x3')(y)
y = BatchNormalization(name='conv5_2_3x3_bn')(y)
y = Conv2D(1024, 1, name='conv5_2_1x1_increase')(y)
y = BatchNormalization(name='conv5_2_1x1_increase_bn')(y)
y = Add(name='conv5_2')([y,y_])
y_ = Activation('relu', name='conv5_2/relu')(y)
y = Conv2D(256, 1, activation='relu', name='conv5_3_1x1_reduce')(y_)
y = BatchNormalization(name='conv5_3_1x1_reduce_bn')(y)
y = ZeroPadding2D(padding=4, name='padding16')(y)
y = Conv2D(256, 3, dilation_rate=4, activation='relu', name='conv5_3_3x3')(y)
y = BatchNormalization(name='conv5_3_3x3_bn')(y)
y = Conv2D(1024, 1, name='conv5_3_1x1_increase')(y)
y = BatchNormalization(name='conv5_3_1x1_increase_bn')(y)
y = Add(name='conv5_3')([y,y_])
y = Activation('relu', name='conv5_3/relu')(y)
h, w = y.shape[1:3].as_list()
pool1 = AveragePooling2D(pool_size=(h,w), strides=(h,w), name='conv5_3_pool1')(y)
pool1 = Lambda(lambda x: tf.image.resize_bilinear(x, size=(h,w)), name='conv5_3_pool1_interp')(pool1)
pool2 = AveragePooling2D(pool_size=(h/2,w/2), strides=(h//2,w//2), name='conv5_3_pool2')(y)
pool2 = Lambda(lambda x: tf.image.resize_bilinear(x, size=(h,w)), name='conv5_3_pool2_interp')(pool2)
pool3 = AveragePooling2D(pool_size=(h/3,w/3), strides=(h//3,w//3), name='conv5_3_pool3')(y)
pool3 = Lambda(lambda x: tf.image.resize_bilinear(x, size=(h,w)), name='conv5_3_pool3_interp')(pool3)
pool6 = AveragePooling2D(pool_size=(h/4,w/4), strides=(h//4,w//4), name='conv5_3_pool6')(y)
pool6 = Lambda(lambda x: tf.image.resize_bilinear(x, size=(h,w)), name='conv5_3_pool6_interp')(pool6)
y = Add(name='conv5_3_sum')([y, pool1, pool2, pool3, pool6])
y = Conv2D(256, 1, activation='relu', name='conv5_4_k1')(y)
y = BatchNormalization(name='conv5_4_k1_bn')(y)
aux_1 = Lambda(lambda x: tf.image.resize_bilinear(x, size=(int(x.shape[1])*2, int(x.shape[2])*2)), name='conv5_4_interp')(y)
y = ZeroPadding2D(padding=2, name='padding17')(aux_1)
y = Conv2D(128, 3, dilation_rate=2, name='conv_sub4')(y)
y = BatchNormalization(name='conv_sub4_bn')(y)
y_ = Conv2D(128, 1, name='conv3_1_sub2_proj')(z)
y_ = BatchNormalization(name='conv3_1_sub2_proj_bn')(y_)
y = Add(name='sub24_sum')([y,y_])
y = Activation('relu', name='sub24_sum/relu')(y)
aux_2 = Lambda(lambda x: tf.image.resize_bilinear(x, size=(int(x.shape[1])*2, int(x.shape[2])*2)), name='sub24_sum_interp')(y)
y = ZeroPadding2D(padding=2, name='padding18')(aux_2)
y_ = Conv2D(128, 3, dilation_rate=2, name='conv_sub2')(y)
y_ = BatchNormalization(name='conv_sub2_bn')(y_)
# (1)
y = Conv2D(32, 3, strides=2, padding='same', activation='relu', name='conv1_sub1')(x)
y = BatchNormalization(name='conv1_sub1_bn')(y)
y = ConvOffset2D(32, name = "conv1_sub1_Offset")(y)
y = Conv2D(32, 3, strides=2, padding='same', activation='relu', name='conv2_sub1')(y)
y = BatchNormalization(name='conv2_sub1_bn')(y)
y = ConvOffset2D(32, name = "conv2_sub1_Offset")(y)
y = Conv2D(64, 3, strides=2, padding='same', activation='relu', name='conv3_sub1')(y)
y = BatchNormalization(name='conv3_sub1_bn')(y)
y = ConvOffset2D(64, name = "conv3_sub1_Offset")(y)
y = Conv2D(128, 1, name='conv3_sub1_proj')(y)
y = BatchNormalization(name='conv3_sub1_proj_bn')(y)
y = Add(name='sub12_sum')([y,y_])
y = Activation('relu', name='sub12_sum/relu')(y)
y = Lambda(lambda x: tf.image.resize_bilinear(x, size=(int(x.shape[1])*2, int(x.shape[2])*2)), name='sub12_sum_interp')(y)
out = Conv2D(n_classes, 1, activation='softmax', name='conv6_cls')(y)
if train:
aux_1 = Conv2D(n_classes, 1, activation='softmax', name='sub4_out')(aux_1)
aux_2 = Conv2D(n_classes, 1, activation='softmax', name='sub24_out')(aux_2)
model = Model(inputs=inp, outputs=[out, aux_2, aux_1])
else:
model = Model(inputs=inp, outputs=out)
if weights_path is not None:
model.load_weights(weights_path, by_name=True)
return model
def get_deform_unet(img_rows,
img_cols,
flt=8,
pool_size=(2, 2, 2),
init_lr=1.0e-5):
print("start building NN")
inputs = Input((img_rows, img_cols, 1))
conv1 = Conv2D(flt, (3, 3), padding='same')(inputs)
conv1 = ConvOffset2D(flt)(conv1)
conv1 = BatchNormalization()(conv1)
conv1 = Activation('relu')(conv1)
conv1 = Conv2D(flt, (3, 3), padding='same')(conv1)
conv1 = ConvOffset2D(flt)(conv1)
conv1 = BatchNormalization()(conv1)
conv1 = Activation('relu')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(flt * 2, (3, 3), padding='same')(pool1)
conv2 = ConvOffset2D(flt * 2)(conv2)
conv2 = BatchNormalization()(conv2)
conv2 = Activation('relu')(conv2)
conv2 = Conv2D(flt * 2, (3, 3), padding='same')(conv2)
conv2 = ConvOffset2D(flt * 2)(conv2)
conv2 = BatchNormalization()(conv2)
conv2 = Activation('relu')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Conv2D(flt * 4, (3, 3), padding='same')(pool2)
conv3 = ConvOffset2D(flt * 4)(conv3)
conv3 = BatchNormalization()(conv3)
conv3 = Activation('relu')(conv3)
conv3 = Conv2D(flt * 4, (3, 3), padding='same')(conv3)
conv3 = ConvOffset2D(flt * 4)(conv3)
conv3 = BatchNormalization()(conv3)
conv3 = Activation('relu')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Conv2D(flt * 8, (3, 3), padding='same')(pool3)
conv4 = ConvOffset2D(flt * 8)(conv4)
conv4 = BatchNormalization()(conv4)
conv4 = Activation('relu')(conv4)
conv4 = Conv2D(flt * 8, (3, 3), padding='same')(conv4)
conv4 = ConvOffset2D(flt * 8)(conv4)
conv4 = BatchNormalization()(conv4)
conv4 = Activation('relu')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Conv2D(flt * 16, (3, 3), padding='same')(pool4)
conv5 = ConvOffset2D(flt * 16)(conv5)
conv5 = BatchNormalization()(conv5)
conv5 = Activation('relu')(conv5)
conv5 = Conv2D(flt * 16, (3, 3), padding='same')(conv5)
conv5 = ConvOffset2D(flt * 16)(conv5)
conv5 = BatchNormalization()(conv5)
conv5 = Activation('relu')(conv5)
up6 = concatenate([
Conv2DTranspose(flt * 16,
(2, 2), strides=(2, 2), padding='same')(conv5), conv4
],
axis=3)
conv6 = Conv2D(flt * 8, (3, 3), padding='same')(up6)
conv6 = ConvOffset2D(flt * 8)(conv6)
conv6 = BatchNormalization()(conv6)
conv6 = Activation('relu')(conv6)
conv6 = Conv2D(flt * 8, (3, 3), padding='same')(conv6)
conv6 = ConvOffset2D(flt * 8)(conv6)
conv6 = BatchNormalization()(conv6)
conv6 = Activation('relu')(conv6)
up7 = concatenate([
Conv2DTranspose(flt * 8,
(2, 2), strides=(2, 2), padding='same')(conv6), conv3
],
axis=3)
conv7 = Conv2D(flt * 4, (3, 3), padding='same')(up7)
conv7 = ConvOffset2D(flt * 4)(conv7)
conv7 = BatchNormalization()(conv7)
conv7 = Activation('relu')(conv7)
conv7 = Conv2D(flt * 4, (3, 3), padding='same')(conv7)
conv7 = ConvOffset2D(flt * 4)(conv7)
conv7 = BatchNormalization()(conv7)
conv7 = Activation('relu')(conv7)
up8 = concatenate([
Conv2DTranspose(flt * 4,
(2, 2), strides=(2, 2), padding='same')(conv7), conv2
],
axis=3)
conv8 = Conv2D(flt * 2, (3, 3), padding='same')(up8)
conv8 = ConvOffset2D(flt * 2)(conv8)
conv8 = BatchNormalization()(conv8)
conv8 = Activation('relu')(conv8)
conv8 = Conv2D(flt * 2, (3, 3), padding='same')(conv8)
conv8 = ConvOffset2D(flt * 2)(conv8)
conv8 = BatchNormalization()(conv8)
conv8 = Activation('relu')(conv8)
up9 = concatenate([
Conv2DTranspose(flt * 2,
(2, 2), strides=(2, 2), padding='same')(conv8), conv1
],
axis=3)
conv9 = Conv2D(flt, (3, 3), padding='same')(up9)
conv9 = ConvOffset2D(flt)(conv9)
conv9 = BatchNormalization()(conv9)
conv9 = Activation('relu')(conv9)
conv9 = Conv2D(flt, (3, 3), padding='same')(conv9)
conv9 = ConvOffset2D(flt)(conv9)
conv9 = BatchNormalization()(conv9)
conv9 = Activation('relu')(conv9)
conv10 = Conv2D(1, (1, 1), activation='sigmoid')(conv9)
model = Model(inputs=[inputs], outputs=[conv10])
model.compile(optimizer=Adam(lr=init_lr),
loss=dice_coef_loss,
metrics=[dice_coef])
return model