问题描述
我正在使用Python 3.7.7。和Tensorflow 2.1.0。
我有一个经过预先训练的VGG16网络，我想获得第一层，即从conv1层到conv5层。
在下图中：
您可以看到卷积编码器-解码器体系结构。我要获取编码器部分，即出现在图像左侧的层：
def vgg16_encoder_decoder(input_size = (200,200,1)):
#################################
# Encoder
#################################
inputs = Input(input_size,name = 'input')
conv1 = Conv2D(64,(3,3),activation = 'relu',padding = 'same',name ='conv1_1')(inputs)
conv1 = Conv2D(64,name ='conv1_2')(conv1)
pool1 = MaxPooling2D(pool_size = (2,2),strides = (2,name = 'pool_1')(conv1)
conv2 = Conv2D(128,name ='conv2_1')(pool1)
conv2 = Conv2D(128,name ='conv2_2')(conv2)
pool2 = MaxPooling2D(pool_size = (2,name = 'pool_2')(conv2)
conv3 = Conv2D(256,name ='conv3_1')(pool2)
conv3 = Conv2D(256,name ='conv3_2')(conv3)
conv3 = Conv2D(256,name ='conv3_3')(conv3)
pool3 = MaxPooling2D(pool_size = (2,name = 'pool_3')(conv3)
conv4 = Conv2D(512,name ='conv4_1')(pool3)
conv4 = Conv2D(512,name ='conv4_2')(conv4)
conv4 = Conv2D(512,name ='conv4_3')(conv4)
pool4 = MaxPooling2D(pool_size = (2,name = 'pool_4')(conv4)
conv5 = Conv2D(512,name ='conv5_1')(pool4)
conv5 = Conv2D(512,name ='conv5_2')(conv5)
conv5 = Conv2D(512,name ='conv5_3')(conv5)
pool5 = MaxPooling2D(pool_size = (2,name = 'pool_5')(conv5)
#################################
# Decoder
#################################
#conv1 = Conv2DTranspose(512,(2,strides = 2,name = 'conv1')(pool5)
upsp1 = UpSampling2D(size = (2,name = 'upsp1')(pool5)
conv6 = Conv2D(512,3,name = 'conv6_1')(upsp1)
conv6 = Conv2D(512,name = 'conv6_2')(conv6)
conv6 = Conv2D(512,name = 'conv6_3')(conv6)
upsp2 = UpSampling2D(size = (2,name = 'upsp2')(conv6)
conv7 = Conv2D(512,name = 'conv7_1')(upsp2)
conv7 = Conv2D(512,name = 'conv7_2')(conv7)
conv7 = Conv2D(512,name = 'conv7_3')(conv7)
zero1 = ZeroPadding2D(padding = ((1,0),(1,0)),data_format = 'channels_last',name='zero1')(conv7)
upsp3 = UpSampling2D(size = (2,name = 'upsp3')(zero1)
conv8 = Conv2D(256,name = 'conv8_1')(upsp3)
conv8 = Conv2D(256,name = 'conv8_2')(conv8)
conv8 = Conv2D(256,name = 'conv8_3')(conv8)
upsp4 = UpSampling2D(size = (2,name = 'upsp4')(conv8)
conv9 = Conv2D(128,name = 'conv9_1')(upsp4)
conv9 = Conv2D(128,name = 'conv9_2')(conv9)
upsp5 = UpSampling2D(size = (2,name = 'upsp5')(conv9)
conv10 = Conv2D(64,name = 'conv10_1')(upsp5)
conv10 = Conv2D(64,name = 'conv10_2')(conv10)
conv11 = Conv2D(1,name = 'conv11')(conv10)
model = Model(inputs = inputs,outputs = conv11,name = 'vgg-16_encoder_decoder')
return model
我训练网络，然后训练它。如何获得编码器零件？换句话说，获得一个模型，该模型仅包含从conv1到pool5的原始图层。
我认为可能是这样的：
model_new = Model(input=model_old.layers[0].input,output=model_old.layers[12].output)
解决方法
from tensorflow.keras.applications import VGG16
from tensorflow.keras.layers import Input,Flatten
from tensorflow.keras import Model
input_shape = (W,H,C)
def encoder(input_shape):
model = VGG16(include_top=False,input_shape=input_shape)
F1 = Flatten()(model.get_layer(index=1).output)
F2 = Flatten()(model.get_layer(index=2).output)
F3 = Flatten()(model.get_layer(index=3).output)
F4 = Flatten()(model.get_layer(index=4).output)
F5 = Flatten()(model.get_layer(index=5).output)
M = Model(model.inputs,[F1,F2,F3,F4,F5])
return M
其中 W，H 图像大小和 C 通道数应等于3。
,要从我的预训练网络中获取伴奏者，我创建了以下功能：
def get_encoder(old_model: Model) -> Model:
# Get encoder
encoder_input: Model = Model(inputs=old_model.layers[0].input,outputs=old_model.layers[14].output)
# Create Global Average Pooling.
encoder_output = GlobalAveragePooling2D()(encoder_input.layers[-1].output)
# Create the encoder adding the GAP layer as output.
encoder: Model = Model(encoder_input.input,encoder_output,name='encoder')
return encoder
重要的是数字14。这是enconder在原始网络中结束的层。顺便说一句，我终于用U-Net代替了VGG-16，所以这个数字仅适用于U-NET 。
通过省略代码的最后20层，我建议以下代码。
model_new = Model(model_old.input,model_old.layers[-20].output) model_new.summary()
如果我错过了解码器最后20层的计数，您可能需要将其稍微调整为-19或-21才能找到最后一个池5。