搭建模型结构代码
import numpy as np
from tensorflow.keras import layers
from tensorflow.keras import models
def VGG19(img_input):
# 512,512,3 -> 512,512,64
x = layers.Conv2D(filters=64, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv1-1')(img_input)
x = layers.Conv2D(filters=64, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv1-2')(x)
conv1 = x
# 512,512,64 -> 256,256,64
x = layers.MaxPooling2D(pool_size=(2,2), strides=(2,2), name='pool1')(x)
# 256,256,64 -> 256,256,128
x = layers.Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv2-1')(x)
x = layers.Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv2-2')(x)
conv2 = x
# 256,256,128 -> 128,128,128
x = layers.MaxPooling2D(pool_size=(2,2), strides=(2,2), name='pool2')(x)
# 128,128,128 -> 128,128,256
x = layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv3-1')(x)
x = layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv3-2')(x)
x = layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv3-3')(x)
x = layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv3-4')(x)
conv3 = x
# 128,128,256 -> 64,64,256
x = layers.MaxPooling2D(pool_size=(2,2), strides=(2,2), name='pool3')(x)
# 64,64,256 -> 64,64,512
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv4-1')(x)
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv4-2')(x)
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv4-3')(x)
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv4-4')(x)
conv4 = x
# 64,64,512 -> 32,32,512
x = layers.MaxPooling2D(pool_size=(2,2), strides=(2,2), name='pool4')(x)
# 32,32,512 -> 32,32,512
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv5-1')(x)
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv5-2')(x)
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv5-3')(x)
x = layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), activation='relu', padding='same', name='conv5-4')(x)
conv5 = x
return conv1, conv2, conv3, conv4, conv5
def Unet(input_shape=(512,512,3), num_classes=21):
'''获取输入的图片的尺寸'''
inputs = layers.Input(input_shape)
'''
获取之前VGG19生成的五个有效特征层
conv1 512,512,64
conv2 256,256,128
conv3 128,128,256
conv4 64,64,512
conv5 32,32,512
'''
conv1, conv2, conv3, conv4, conv5 = VGG19(inputs)
# 32,32,512 -> 64,64,512
U5_up = layers.UpSampling2D(size=(2,2), interpolation='nearest')(conv5)
# 64,64,512 + 64,64,512 -> 64,64,1024
U4 = layers.Concatenate(axis=3)([conv4, U5_up])
# 64,64,1024 -> 64,64,512
U4 = layers.Conv2D(filters=512, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U4)
U4 = layers.Conv2D(filters=512, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U4)
# 64,64,512 -> 128,128,512
U4_up = layers.UpSampling2D(size=(2,2), interpolation='nearest')(U4)
# 128,128,512 + 128,128,256 -> 128,128,768
U3 = layers.Concatenate(axis=3)([conv3, U4_up])
# 128,128,768 -> 128,128,256
U3 = layers.Conv2D(filters=256, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U3)
U3 = layers.Conv2D(filters=256, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U3)
# 128,128,256 -> 256,256,256
U3_up = layers.UpSampling2D(size=(2,2), interpolation='nearest')(U3)
# 256,256,256 + 256,256,128 -> 256,256,384
U2 = layers.Concatenate(axis=3)([conv2, U3_up])
# 256,256,384 -> 256,256,128
U2 = layers.Conv2D(filters=128, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U2)
U2 = layers.Conv2D(filters=128, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U2)
# 256,256,128 -> 512,512,128
U2_up = layers.UpSampling2D(size=(2,2), interpolation='nearest')(U2)
# 512,512,128 + 512,512,64 -> 512,512,192
U1 = layers.Concatenate(axis=3)([conv1, U2_up])
# 512,512,192 -> 512,512,64
U1 = layers.Conv2D(filters=64, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U1)
U1 = layers.Conv2D(filters=64, kernel_size=3, strides=1, activation='relu', padding='same', kernel_initializer='he_normal')(U1)
# 512,512,64 -> 512,512,num_classes
outputs = layers.Conv2D(filters=num_classes, kernel_size=1, strides=1, activation='softmax')(U1)
model = models.Model(inputs=inputs, outputs=outputs)
return model
测试模型并的出模型结构层次
image_size = (512,512,3)
num_classes = 21
model = Unet(image_size, num_classes)
model.summary()
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 512, 512, 3) 0
__________________________________________________________________________________________________
conv1-1 (Conv2D) (None, 512, 512, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv1-2 (Conv2D) (None, 512, 512, 64) 36928 conv1-1[0][0]
__________________________________________________________________________________________________
pool1 (MaxPooling2D) (None, 256, 256, 64) 0 conv1-2[0][0]
__________________________________________________________________________________________________
conv2-1 (Conv2D) (None, 256, 256, 128 73856 pool1[0][0]
__________________________________________________________________________________________________
conv2-2 (Conv2D) (None, 256, 256, 128 147584 conv2-1[0][0]
__________________________________________________________________________________________________
pool2 (MaxPooling2D) (None, 128, 128, 128 0 conv2-2[0][0]
__________________________________________________________________________________________________
conv3-1 (Conv2D) (None, 128, 128, 256 295168 pool2[0][0]
__________________________________________________________________________________________________
conv3-2 (Conv2D) (None, 128, 128, 256 590080 conv3-1[0][0]
__________________________________________________________________________________________________
conv3-3 (Conv2D) (None, 128, 128, 256 590080 conv3-2[0][0]
__________________________________________________________________________________________________
conv3-4 (Conv2D) (None, 128, 128, 256 590080 conv3-3[0][0]
__________________________________________________________________________________________________
pool3 (MaxPooling2D) (None, 64, 64, 256) 0 conv3-4[0][0]
__________________________________________________________________________________________________
conv4-1 (Conv2D) (None, 64, 64, 512) 1180160 pool3[0][0]
__________________________________________________________________________________________________
conv4-2 (Conv2D) (None, 64, 64, 512) 2359808 conv4-1[0][0]
__________________________________________________________________________________________________
conv4-3 (Conv2D) (None, 64, 64, 512) 2359808 conv4-2[0][0]
__________________________________________________________________________________________________
conv4-4 (Conv2D) (None, 64, 64, 512) 2359808 conv4-3[0][0]
__________________________________________________________________________________________________
pool4 (MaxPooling2D) (None, 32, 32, 512) 0 conv4-4[0][0]
__________________________________________________________________________________________________
conv5-1 (Conv2D) (None, 32, 32, 512) 2359808 pool4[0][0]
__________________________________________________________________________________________________
conv5-2 (Conv2D) (None, 32, 32, 512) 2359808 conv5-1[0][0]
__________________________________________________________________________________________________
conv5-3 (Conv2D) (None, 32, 32, 512) 2359808 conv5-2[0][0]
__________________________________________________________________________________________________
conv5-4 (Conv2D) (None, 32, 32, 512) 2359808 conv5-3[0][0]
__________________________________________________________________________________________________
up_sampling2d (UpSampling2D) (None, 64, 64, 512) 0 conv5-4[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 64, 64, 1024) 0 conv4-4[0][0]
up_sampling2d[0][0]
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 64, 64, 512) 4719104 concatenate[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 64, 64, 512) 2359808 conv2d[0][0]
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D) (None, 128, 128, 512 0 conv2d_1[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 128, 128, 768 0 conv3-4[0][0]
up_sampling2d_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 128, 128, 256 1769728 concatenate_1[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 128, 128, 256 590080 conv2d_2[0][0]
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D) (None, 256, 256, 256 0 conv2d_3[0][0]
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 256, 256, 384 0 conv2-2[0][0]
up_sampling2d_2[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 256, 256, 128 442496 concatenate_2[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 256, 256, 128 147584 conv2d_4[0][0]
__________________________________________________________________________________________________
up_sampling2d_3 (UpSampling2D) (None, 512, 512, 128 0 conv2d_5[0][0]
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 512, 512, 192 0 conv1-2[0][0]
up_sampling2d_3[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 512, 512, 64) 110656 concatenate_3[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 512, 512, 64) 36928 conv2d_6[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 512, 512, 21) 1365 conv2d_7[0][0]
==================================================================================================
Total params: 30,202,133
Trainable params: 30,202,133
Non-trainable params: 0
__________________________________________________________________________________________________