def getResidualBlock(I, mode, filter_size, featmaps, activation, dropout,
shortcut, convArgs, bnArgs):
"""Get residual block."""
if mode == "real":
O = BatchNormalization(**bnArgs)(I)
elif mode == "complex":
O = ComplexBatchNormalization(**bnArgs)(I)
elif mode == "quaternion":
O = QuaternionBatchNormalization(**bnArgs)(I)
O = Activation(activation)(O)
if shortcut == 'regular':
if mode == "real":
O = Conv2D(featmaps, filter_size, **convArgs)(O)
elif mode == "complex":
O = ComplexConv2D(featmaps, filter_size, **convArgs)(O)
elif mode == "quaternion":
O = QuaternionConv2D(featmaps, filter_size, **convArgs)(O)
elif shortcut == 'projection':
if mode == "real":
O = Conv2D(featmaps, filter_size, **convArgs)(O)
elif mode == "complex":
O = ComplexConv2D(featmaps, filter_size, **convArgs)(O)
elif mode == "quaternion":
O = QuaternionConv2D(featmaps, filter_size, **convArgs)(O)
if mode == "real":
O = BatchNormalization(**bnArgs)(O)
O = Activation(activation)(O)
O = Conv2D(featmaps, filter_size, **convArgs)(O)
elif mode == "complex":
O = ComplexBatchNormalization(**bnArgs)(O)
O = Activation(activation)(O)
O = ComplexConv2D(featmaps, filter_size, **convArgs)(O)
elif mode == "quaternion":
O = QuaternionBatchNormalization(**bnArgs)(O)
O = Activation(activation)(O)
O = QuaternionConv2D(featmaps, filter_size, **convArgs)(O)
if shortcut == 'regular':
O = Add()([O, I])
elif shortcut == 'projection':
if mode == "real":
X = Conv2D(featmaps, (1, 1), **convArgs)(I)
O = Concatenate(1)([X, O])
elif mode == "complex":
X = ComplexConv2D(featmaps, (1, 1), **convArgs)(I)
O_real = Concatenate(1)([GetReal()(X), GetReal()(O)])
O_imag = Concatenate(1)([GetImag()(X), GetImag()(O)])
O = Concatenate(1)([O_real, O_imag])
elif mode == "quaternion":
X = QuaternionConv2D(featmaps, (1, 1), **convArgs)(I)
O_r = Concatenate(1)([GetR()(X), GetR()(O)])
O_i = Concatenate(1)([GetI()(X), GetI()(O)])
O_j = Concatenate(1)([GetJ()(X), GetJ()(O)])
O_k = Concatenate(1)([GetK()(X), GetK()(O)])
O = Concatenate(1)([O_r, O_i, O_j, O_k])
return O
def getModel(params):
mode = params.mode
n = params.num_blocks
sf = params.start_filter
activation = params.act
dropout = params.dropout
inputShape = (3, 93, 310)
channelAxis = 1
filsize = (3, 3)
convArgs = {
"padding": "same",
"use_bias": False,
"kernel_regularizer": l2(0.0001),
}
bnArgs = {
"axis": channelAxis,
"momentum": 0.9,
"epsilon": 1e-04,
"scale": False
}
convArgs.update({"kernel_initializer": params.init})
# Create the vector channels
R = Input(shape=inputShape)
if mode != "quaternion":
I = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
O = concatenate([R, I], axis=channelAxis)
else:
I = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
J = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
K = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
O = concatenate([R, I, J, K], axis=channelAxis)
if mode == "real":
O = Conv2D(sf, filsize, **convArgs)(O)
O = BatchNormalization(**bnArgs)(O)
elif mode == "complex":
O = ComplexConv2D(sf, filsize, **convArgs)(O)
O = ComplexBatchNormalization(**bnArgs)(O)
else:
O = QuaternionConv2D(sf, filsize, **convArgs)(O)
O = QuaternionBatchNormalization(**bnArgs)(O)
O = Activation(activation)(O)
for i in range(n):
O = getResidualBlock(O, mode, filsize, sf, activation, dropout, 'regular', convArgs, bnArgs)
O = getResidualBlock(O, mode, filsize, sf, activation, dropout, 'projection', convArgs, bnArgs)
for i in range(n-1):
O = getResidualBlock(O, mode, filsize, sf*2, activation, dropout, 'regular', convArgs, bnArgs)
O = getResidualBlock(O, mode, filsize, sf*2, activation, dropout, 'projection', convArgs, bnArgs)
for i in range(n-1):
O = getResidualBlock(O, mode, filsize, sf*4, activation, dropout, 'regular', convArgs, bnArgs)
# heatmap output
O = Convolution2D(1, 1, activation='sigmoid')(O)
model = Model(R, O)
opt = SGD (lr = params.lr,
momentum = params.momentum,
decay = params.decay,
nesterov = True,
clipnorm = params.clipnorm)
model.compile(opt, dice_coef_loss)
return model
datagen.fit(x_train)
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
#将灰度图片转为四自由度图片
R = Input(shape=input_shape)
I = learnVectorBlock(R)
J = learnVectorBlock(R)
K = learnVectorBlock(R)
O = concatenate([R, I, J, K], axis=-1)
O = QuaternionConv2D(64, (5, 5),
activation='relu',
padding="same",
kernel_initializer='quaternion')(O)
O = QuaternionConv2D(64, (5, 5),
activation='relu',
padding="same",
kernel_initializer='quaternion')(O)
O = QuaternionConv2D(32, (5, 5),
activation='relu',
padding="same",
kernel_initializer='quaternion')(O)
# O = Convolution2D(256, (5, 5), activation='relu', padding="same")(R)
# O = Convolution2D(256, (5, 5), activation='relu', padding="same")(O)
# O = Convolution2D(128, (5, 5), activation='relu', padding="same")(O)
O = Flatten()(O)
def getModel(params):
mode = params.mode
n = params.num_blocks
sf = params.start_filter
dataset = params.dataset
activation = params.act
inputShape = (3, 512, 512)
channelAxis = 1
filsize = (3, 3)
convArgs = {
"padding": "same",
"use_bias": False,
"kernel_regularizer": l2(0.0001),
}
bnArgs = {"axis": channelAxis, "momentum": 0.9, "epsilon": 1e-04}
convArgs.update({"kernel_initializer": params.init})
# Create the vector channels
R = Input(shape=inputShape)
if mode != "quaternion":
I = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
O = concatenate([R, I], axis=channelAxis)
else:
I = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
J = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
K = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
O = concatenate([R, I, J, K], axis=channelAxis)
if mode == "real":
O = Conv2D(sf, filsize, **convArgs)(O)
O = BatchNormalization(**bnArgs)(O)
elif mode == "complex":
O = ComplexConv2D(sf, filsize, **convArgs)(O)
O = ComplexBatchNormalization(**bnArgs)(O)
else:
O = QuaternionConv2D(64, (5, 5), **convArgs)(O)
O = QuaternionBatchNormalization(**bnArgs)(O)
O = Activation(activation)(O)
O = MaxPooling2D(pool_size=(3, 3))
# conv2
O = QuaternionConv2D(128, (5, 5), **convArgs)
O = Activation(activation)(O)
O = MaxPooling2D(pool_size=(3, 3))
#conv3
O = QuaternionConv2D(256, (3, 3), **convArgs)
O = Activation(activation)(O)
O = MaxPooling2D(pool_size=(3, 3))
#conv4
O = QuaternionConv2D(512, (3, 3), **convArgs)
O = Activation(activation)(O)
O = MaxPooling2D(pool_size=(3, 3))
#conv5
O = QuaternionConv2D(16, (3, 3), **convArgs)
O = Activation(activation)(O)
O = MaxPooling2D(pool_size=(3, 3))
# for i in range(n):
# O = getResidualBlock(O, mode, filsize, sf, activation, 'regular', convArgs, bnArgs)
#
# O = getResidualBlock(O, mode, filsize, sf, activation, 'projection', convArgs, bnArgs)
#
# for i in range(n - 1):
# O = getResidualBlock(O, mode, filsize, sf * 2, activation, 'regular', convArgs, bnArgs)
#
# O = getResidualBlock(O, mode, filsize, sf * 2, activation, 'projection', convArgs, bnArgs)
#
# for i in range(n - 1):
# O = getResidualBlock(O, mode, filsize, sf * 4, activation, 'regular', convArgs, bnArgs)
# O = AveragePooling2D(pool_size=(8, 8))(O)
# Flatten
O = Flatten()(O)
# Dense
O = Dense(1000, activation='relu', kernel_regularizer=l2(0.0001))(O)
model = Model(R, O)
return model
def getModel(params):
mode = params.mode
n = params.num_blocks
sf = params.start_filter
dataset = params.dataset
activation = params.act
inputShape = (3, 32, 32)
channelAxis = 1
filsize = (3, 3)
convArgs = {
"padding": "same",
"use_bias": False,
"kernel_regularizer": l2(0.0001),
}
bnArgs = {
"axis": channelAxis,
"momentum": 0.9,
"epsilon": 1e-04
}
convArgs.update({"kernel_initializer": params.init})
# Create the vector channels
R = Input(shape=inputShape)
if mode != "quaternion":
I = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
O = concatenate([R, I], axis=channelAxis)
else:
I = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
J = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
K = learnVectorBlock(R, 3, filsize, 'relu', bnArgs)
O = concatenate([R, I, J, K], axis=channelAxis)
if mode == "real":
O = Conv2D(sf, filsize, **convArgs)(O)
O = BatchNormalization(**bnArgs)(O)
elif mode == "complex":
O = ComplexConv2D(sf, filsize, **convArgs)(O)
O = ComplexBatchNormalization(**bnArgs)(O)
else:
O = QuaternionConv2D(sf, filsize, **convArgs)(O)
O = QuaternionBatchNormalization(**bnArgs)(O)
O = Activation(activation)(O)
for i in range(n):
O = getResidualBlock(O, mode, filsize, sf, activation, 'regular', convArgs, bnArgs)
O = getResidualBlock(O, mode, filsize, sf, activation, 'projection', convArgs, bnArgs)
for i in range(n-1):
O = getResidualBlock(O, mode, filsize, sf*2, activation, 'regular', convArgs, bnArgs)
O = getResidualBlock(O, mode, filsize, sf*2, activation, 'projection', convArgs, bnArgs)
for i in range(n-1):
O = getResidualBlock(O, mode, filsize, sf*4, activation, 'regular', convArgs, bnArgs)
O = AveragePooling2D(pool_size=(8, 8))(O)
# Flatten
O = Flatten()(O)
# Dense
if dataset == 'cifar10':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'cifar100':
O = Dense(100, activation='softmax', kernel_regularizer=l2(0.0001))(O)
model = Model(R, O)
opt = SGD (lr = params.lr,
momentum = params.momentum,
decay = params.decay,
nesterov = True,
clipnorm = params.clipnorm)
model.compile(opt, 'categorical_crossentropy', metrics=['accuracy'])
return model