def xceptionv3_exit(self, sc1, sc2, x):
"""
:param sc1:
:param sc2:
:param x:
:return:
"""
sc_1 = keras.layers.SeparableConv2D(filters=sc1,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)
ac1 = Mish()
bn1 = keras.layers.BatchNormalization()
sc_2 = keras.layers.SeparableConv2D(filters=sc2,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)
ac2 = Mish()
bn2 = keras.layers.BatchNormalization()
# maxpool 3×3
mp_3 = keras.layers.MaxPool2D(pool_size=3, strides=2, padding='same')
shortcut_conv = keras.layers.Conv2D(filters=sc2,
kernel_size=1,
strides=2,
padding='same')
p1 = mp_3(ac2(bn2(sc_2(ac1(bn1(sc_1(x)))))))
return Mish()(p1 + shortcut_conv(x))
def xceptionv3_middle(self, x):
sc_1 = keras.layers.SeparableConv2D(filters=728,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)
ac1 = Mish()
bn1 = keras.layers.BatchNormalization()
sc_2 = keras.layers.SeparableConv2D(filters=728,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)
ac2 = Mish()
bn2 = keras.layers.BatchNormalization()
sc_3 = keras.layers.SeparableConv2D(filters=728,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)
p1 = sc_3(ac2(bn2(sc_2(ac1(bn1(sc_1(x)))))))
return Mish()(p1 + x)
def InesNet(c_out):
initial = nn.Sequential(NiceDownscale(3,64,k=6,activation=False))
blockA = nn.Sequential(concat_downscale(64,128),
ResBlock(128),
NiceDownscale(128,256),
ResBlock(256,g=2),
ResBlock(256,bottle=192),
)
blockB = nn.Sequential(concat_downscale(256,512),
ResBlock(512,bottle=256,g=4),
ResBlock(512,bottle=384,g=2),
ResBlock(512,g=4),
ResBlock(512,g=2),
concat_downscale(512,768),
ResBlock(768,bottle=512,g=4),
ResBlock(768,g=2),
ResBlock(768,bottle=384,g=3),
ResBlock(768))
classifier = nn.Sequential(nn.AdaptiveAvgPool2d(1),
fv.Flatten(),
Mish(),
nn.BatchNorm1d(768),
nn.Linear(768,c_out))
return nn.Sequential(initial, blockA, blockB, classifier)
def alexnet(self):
"""
定义AlexNet的结构
:return:
"""
output_num = 3
net = keras.models.Sequential()
net.add(
keras.layers.Conv2D(filters=96,
kernel_size=11,
strides=4,
padding='same',
input_shape=[width, height, channels]))
net.add(keras.layers.BatchNormalization())
net.add(Mish())
net.add(keras.layers.MaxPool2D(pool_size=3, strides=2))
net.add(
keras.layers.Conv2D(filters=256,
kernel_size=5,
strides=1,
padding='same'))
net.add(keras.layers.BatchNormalization())
net.add(Mish())
net.add(keras.layers.MaxPool2D(pool_size=3, strides=2))
net.add(keras.layers.Conv2D(filters=384, kernel_size=3,
padding='same'))
net.add(keras.layers.BatchNormalization())
net.add(Mish())
net.add(keras.layers.Conv2D(filters=384, kernel_size=3,
padding='same'))
net.add(keras.layers.BatchNormalization())
net.add(Mish())
net.add(keras.layers.Conv2D(filters=256, kernel_size=3,
padding='same'))
net.add(keras.layers.BatchNormalization())
net.add(Mish())
net.add(keras.layers.MaxPool2D(pool_size=2, strides=2))
net.add(keras.layers.Flatten())
net.add(keras.layers.Dense(1024))
net.add(Mish())
net.add(keras.layers.Dropout(0.5))
net.add(keras.layers.Dense(512))
net.add(Mish())
net.add(keras.layers.Dropout(0.5))
net.add(keras.layers.Dense(output_num, activation=r'softmax'))
net.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
net.summary()
return net
def xception(self):
output_num = 3
inpt = keras.layers.Input(shape=[width, height, channels])
# block 1
x = keras.layers.Conv2D(filters=32,
kernel_size=3,
strides=2,
padding='valid')(inpt)
x = keras.layers.BatchNormalization()(x)
x = Mish()(x)
x = keras.layers.Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = Mish()(x)
# block 2
x = self.xceptionv3_entry(sc1=128, sc2=128, x=x)
# block 3
x = self.xceptionv3_entry(sc1=256, sc2=256, x=x)
# block 4
x = self.xceptionv3_entry(sc1=728, sc2=728, x=x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_middle(x)
x = self.xceptionv3_exit(sc1=728, sc2=1024, x=x)
x = keras.layers.SeparableConv2D(filters=1536,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)(x)
x = keras.layers.BatchNormalization()(x)
x = Mish()(x)
x = keras.layers.SeparableConv2D(filters=2048,
kernel_size=3,
strides=1,
padding='same',
depth_multiplier=1)(x)
x = keras.layers.BatchNormalization()(x)
x = Mish()(x)
x = keras.layers.GlobalAveragePooling2D()(x)
# 6 layers
x = keras.layers.Flatten()(x)
x = keras.layers.Dropout(0.5)(x)
x = keras.layers.Dense(output_num, activation=r'softmax')(x)
model = keras.Model(inputs=inpt, outputs=x)
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
return model
def inceptionv2(self):
output_num = 3
inpt = keras.layers.Input(shape=[width, height, channels])
# 1 layers
x = keras.layers.SeparableConv2D(filters=64,
kernel_size=7,
strides=2,
padding='same')(inpt)
x = keras.layers.BatchNormalization()(x)
x = Mish()(x)
x = keras.layers.MaxPool2D(pool_size=3, strides=2, padding='same')(x)
# 2 layers
x = keras.layers.Conv2D(filters=64,
kernel_size=1,
strides=1,
padding='same')(x)
x = Mish()(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Conv2D(filters=192,
kernel_size=3,
strides=1,
padding='same')(x)
x = Mish()(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPool2D(pool_size=3, strides=2, padding='same')(x)
# 3 layers
x = self.inceptionv2_module_v1(n1_1=64,
n2_1=64,
n2_3=64,
n3_1=64,
n3_5=96,
n4_1=32,
x=x)
x = self.inceptionv2_module_v1(n1_1=64,
n2_1=64,
n2_3=96,
n3_1=64,
n3_5=96,
n4_1=64,
x=x)
x = self.inceptionv2_module_v1(n1_1=0,
n2_1=128,
n2_3=160,
n3_1=64,
n3_5=96,
n4_1=0,
x=x,
is_max=True,
strides=2)
# 4 layers
x = self.inceptionv2_module_v1(n1_1=224,
n2_1=64,
n2_3=96,
n3_1=96,
n3_5=128,
n4_1=128,
x=x)
x = self.inceptionv2_module_v1(n1_1=192,
n2_1=96,
n2_3=128,
n3_1=96,
n3_5=128,
n4_1=128,
x=x)
x = self.inceptionv2_module_v1(n1_1=128,
n2_1=128,
n2_3=160,
n3_1=128,
n3_5=160,
n4_1=128,
x=x)
x = self.inceptionv2_module_v1(n1_1=96,
n2_1=128,
n2_3=160,
n3_1=160,
n3_5=192,
n4_1=128,
x=x)
x = self.inceptionv2_module_v1(n1_1=0,
n2_1=128,
n2_3=192,
n3_1=192,
n3_5=256,
n4_1=0,
x=x,
is_max=True,
strides=2)
# 5 layers
x = self.inceptionv2_module_v1(n1_1=352,
n2_1=192,
n2_3=320,
n3_1=160,
n3_5=224,
n4_1=128,
x=x)
x = self.inceptionv2_module_v1(n1_1=352,
n2_1=192,
n2_3=320,
n3_1=192,
n3_5=224,
n4_1=128,
x=x,
is_max=True)
x = keras.layers.AvgPool2D(pool_size=7, strides=7, padding='same')(x)
# 6 layers
x = keras.layers.Flatten()(x)
x = keras.layers.Dropout(0.7)(x)
x = keras.layers.Dense(output_num, activation=r'softmax')(x)
model = keras.Model(inputs=inpt, outputs=x)
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
return model
def inceptionv2_module_v1(self,
n1_1,
n2_1,
n2_3,
n3_1,
n3_5,
n4_1,
x,
is_max=False,
strides=1):
p1_conv_1 = keras.layers.Conv2D(filters=n1_1,
kernel_size=1,
strides=strides,
padding='same')
p1_ac1 = Mish()
p2_conv_1 = keras.layers.Conv2D(filters=n2_1,
kernel_size=1,
strides=1,
padding='same')
p2_ac1 = Mish()
p2_conv_3 = keras.layers.Conv2D(filters=n2_3,
kernel_size=3,
strides=strides,
padding='same')
p2_ac2 = Mish()
p3_conv_1 = keras.layers.Conv2D(filters=n3_1,
kernel_size=1,
strides=strides,
padding='same')
p3_ac1 = Mish()
p3_conv_3_1 = keras.layers.Conv2D(filters=n3_5,
kernel_size=3,
strides=1,
padding='same')
p3_ac2 = Mish()
p3_conv_3_2 = keras.layers.Conv2D(filters=n3_5,
kernel_size=3,
strides=1,
padding='same')
p3_ac3 = Mish()
if is_max:
p4_maxpool_3 = keras.layers.MaxPool2D(pool_size=3,
strides=strides,
padding='same')
else:
p4_maxpool_3 = keras.layers.AvgPool2D(pool_size=3,
strides=strides,
padding='same')
p4_conv_1 = keras.layers.Conv2D(filters=n4_1,
kernel_size=1,
strides=1,
padding='same')
p4_ac1 = Mish()
p1 = p1_ac1(p1_conv_1(x))
p2 = p2_ac2(p2_conv_3(p2_ac1(p2_conv_1(x))))
p3 = p3_ac3(p3_conv_3_2(p3_ac2(p3_conv_3_1(p3_ac1(p3_conv_1(x))))))
if n4_1 > 0:
p4 = p4_ac1(p4_conv_1(p4_maxpool_3(x)))
else:
p4 = p4_maxpool_3(x)
if n1_1 > 0:
return keras.layers.concatenate([p1, p2, p3, p4], 3)
else:
return keras.layers.concatenate([p2, p3, p4], 3)
x = tf.add(x,c3) x = PointPooling(batch_sample_xyz=xyz128,sampling=xyz32, poolN=9)(x) x = BatchNormalization()(x) c4 = Conv1D(1024,1)(x) x = FPAC_Layer(xyz = xyz32, cin = 512, cout = 1024, m1=[3,9,1], m2=[524288,64,32], mr=[16384,64], mid=32, maxn=32, framepoints=framepoints4,numframe=num_framepoint,N=2048, l2=l2, dtype=dtype)(x) x = tf.add(x,c4) x = PointPooling(batch_sample_xyz=xyz32,sampling=xyz8,poolN=8)(x) x = BatchNormalization()(x) x = MaxPooling1D(pool_size=8)(x) x = Dense(512)(x) x = Mish()(x) x = BatchNormalization()(x) x = Dropout(rate=0.3)(x) x = Dense(256)(x) x = Mish()(x) x = BatchNormalization()(x) x = Dropout(rate=0.3)(x) x = Dense(128)(x) x = Mish()(x) x = BatchNormalization()(x) x = Dropout(rate=0.3)(x) x = Dense(40, activation = 'softmax')(x) prediction = Flatten()(x) model = Model(inputs=[inputs,idx512,idx128,idx32,idx8], outputs=prediction)