【发布时间】：2018-10-09 22:31:30
【问题描述】：
我想在弹出最后一层然后添加 2 个 FC(Dense) 层后使用预训练的 vgg-19。我想为这些 FC 层之一提供一个二进制变量（男性或女性）的输入。我怎样才能实现它。
【问题讨论】：
标签： keras keras-layer
【发布时间】：2018-10-09 22:31:30
【问题描述】：
您需要一个双输入单输出网络，其中每个输入都有自己的特征提取，并且在预测最终输出之前融合了两个特征。
下面是一个例子。
import keras
from keras.applications.vgg19 import VGG19
from keras.layers import Input, Dense, Concatenate
from keras.models import Model
#-------------------------------------------------------------------------------
# Define your new inputs
# Here I pretend that your new task is a classification task over 100 classes
#-------------------------------------------------------------------------------
gender_input = Input(shape=(1,), name='gender_input')
image_input = Input(shape=(224,224,3), name='image_input')
num_classes = 100
#-------------------------------------------------------------------------------
# define your pretrained feature extraction
# you may do something different than below
# but the point here is to have a pretrained model for `featex`
# so you may define it differently
#-------------------------------------------------------------------------------
pretrained = VGG19()
featex = Model( pretrained.input, pretrained.layers[-2].output, name='vgg_pop_last' )
# consider to freeze all weights in featex to speed-up training
image_feat = featex( image_input )
#-------------------------------------------------------------------------------
# From here, you may play with different network architectures
# Below is just one example
#-------------------------------------------------------------------------------
image_feat = Dense(128, activation='relu', name='image_fc')( image_feat )
gender_feat = Dense(16, activation='relu', name='gender_fc')( gender_input )
# mix information from both information
# note: concatenation is only one of plausible way to mix information
concat_feat = Concatenate(axis=-1,name='concat_fc')([image_feat, gender_feat])
# perform final prediction
target = Dense(num_classes, activation='softmax', name='pred_class')( concat_feat )
#-------------------------------------------------------------------------------
# Here is your new model which contains two inputs and one new target
#-------------------------------------------------------------------------------
model = Model( inputs=[image_input, gender_input], outputs=target, name='myModel')
print model.summary()
【讨论】：
如果你想重用第5层的old_model的层作为新模型输出，你需要像这样重建模型，关键是定义一个输入，并将其传递到第 5 层。它可能看起来像这样：
FC = ... # your FC layers
old_model = models.load_model(MODEL_DIR)
old_model.layers.pop(-1) # Get rid of the classification layer
for i in range(4): # Get rid of the first 4 layer
old_model.layers.pop(0) # care pop index while iterating
input = Input(shape=(INPUT_DIM,))
new_model = Model(inputs=input, outputs=FC(old_model(input)))
【讨论】：