如何從經過訓練的 VGG16 網絡中獲取編碼器

Question

我正在使用 Python 3.7.7。 和 Tensorflow 2.1.0。

我有一個預訓練的 VGG16 網絡，我想獲得第一層，即從 conv1 層到 conv5 層。

在下圖中：

您可以看到卷積編碼器-解碼器架構。 我想得到編碼器部分，也就是出現在圖像左側的圖層：

這只是一個例子，但如果我從這個 function 獲得 VGG16：

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, (3, 3), activation = 'relu', padding = 'same', name ='conv1_2')(conv1)
    pool1 = MaxPooling2D(pool_size = (2,2), strides = (2,2), name = 'pool_1')(conv1)

    conv2 = Conv2D(128, (3, 3), activation = 'relu', padding = 'same', name ='conv2_1')(pool1)
    conv2 = Conv2D(128, (3, 3), activation = 'relu', padding = 'same', name ='conv2_2')(conv2)
    pool2 = MaxPooling2D(pool_size = (2,2), strides = (2,2), name = 'pool_2')(conv2)
    
    conv3 = Conv2D(256, (3, 3), activation = 'relu', padding = 'same', name ='conv3_1')(pool2)
    conv3 = Conv2D(256, (3, 3), activation = 'relu', padding = 'same', name ='conv3_2')(conv3)
    conv3 = Conv2D(256, (3, 3), activation = 'relu', padding = 'same', name ='conv3_3')(conv3)
    pool3 = MaxPooling2D(pool_size = (2,2), strides = (2,2), name = 'pool_3')(conv3)
    
    conv4 = Conv2D(512, (3, 3), activation = 'relu', padding = 'same', name ='conv4_1')(pool3)
    conv4 = Conv2D(512, (3, 3), activation = 'relu', padding = 'same', name ='conv4_2')(conv4)
    conv4 = Conv2D(512, (3, 3), activation = 'relu', padding = 'same', name ='conv4_3')(conv4)
    pool4 = MaxPooling2D(pool_size = (2,2), strides = (2,2), name = 'pool_4')(conv4)

    conv5 = Conv2D(512, (3, 3), activation = 'relu', padding = 'same', name ='conv5_1')(pool4)
    conv5 = Conv2D(512, (3, 3), activation = 'relu', padding = 'same', name ='conv5_2')(conv5)
    conv5 = Conv2D(512, (3, 3), activation = 'relu', padding = 'same', name ='conv5_3')(conv5)
    pool5 = MaxPooling2D(pool_size = (2,2), strides = (2,2), name = 'pool_5')(conv5)

    #################################
    # Decoder
    #################################
    #conv1 = Conv2DTranspose(512, (2, 2), strides = 2, name = 'conv1')(pool5)

    upsp1 = UpSampling2D(size = (2,2), name = 'upsp1')(pool5)
    conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', name = 'conv6_1')(upsp1)
    conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', name = 'conv6_2')(conv6)
    conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', name = 'conv6_3')(conv6)

    upsp2 = UpSampling2D(size = (2,2), name = 'upsp2')(conv6)
    conv7 = Conv2D(512, 3, activation = 'relu', padding = 'same', name = 'conv7_1')(upsp2)
    conv7 = Conv2D(512, 3, activation = 'relu', padding = 'same', name = 'conv7_2')(conv7)
    conv7 = Conv2D(512, 3, activation = 'relu', padding = 'same', name = 'conv7_3')(conv7)
    zero1 = ZeroPadding2D(padding =  ((1, 0), (1, 0)), data_format = 'channels_last', name='zero1')(conv7)

    upsp3 = UpSampling2D(size = (2,2), name = 'upsp3')(zero1)
    conv8 = Conv2D(256, 3, activation = 'relu', padding = 'same', name = 'conv8_1')(upsp3)
    conv8 = Conv2D(256, 3, activation = 'relu', padding = 'same', name = 'conv8_2')(conv8)
    conv8 = Conv2D(256, 3, activation = 'relu', padding = 'same', name = 'conv8_3')(conv8)

    upsp4 = UpSampling2D(size = (2,2), name = 'upsp4')(conv8)
    conv9 = Conv2D(128, 3, activation = 'relu', padding = 'same', name = 'conv9_1')(upsp4)
    conv9 = Conv2D(128, 3, activation = 'relu', padding = 'same', name = 'conv9_2')(conv9)

    upsp5 = UpSampling2D(size = (2,2), name = 'upsp5')(conv9)
    conv10 = Conv2D(64, 3, activation = 'relu', padding = 'same', name = 'conv10_1')(upsp5)
    conv10 = Conv2D(64, 3, activation = 'relu', padding = 'same', name = 'conv10_2')(conv10)

    conv11 = Conv2D(1, 3, activation = 'relu', padding = 'same', name = 'conv11')(conv10)

    model = Model(inputs = inputs, outputs = conv11, name = 'vgg-16_encoder_decoder')

    return model

我訓練網絡，然后訓練它。 如何獲得編碼器部分？ 換句話說，得到一個 model ，只有從conv1到pool5的原始層。

我認為它可能是這樣的：

model_new = Model(input=model_old.layers[0].input, 
                  output=model_old.layers[12].output)

Answer 1

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是圖像大小， C 通道數應該等於 3。

Answer 2

為了從我預訓練的網絡中獲取編碼，我創建了這個 function：

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

重要的是數字14 。 這是編碼器在原始網絡中結束的層。 順便說一句，我終於用U-Net代替了VGG-16 ，所以這個數字只適用於 U-NET 。

Answer 3

我建議使用以下代碼，從您的代碼中省略最后 20 層。

model_new = Model(model_old.input, model_old.layers[-20].output) model_new.summary()

如果我錯過了解碼器最后 20 層的計數，您可能需要將其稍微調整為 -19 或 -21 以找到最后一個 pool5。