Keras使用Pretrain自己训练好的模型，迁移学习1（tensorflow相关）

目的

使用之前在其他场景下训练好的CNN模型，用于当前图像的识别分类。已训练好的模型结构如下图，我只需要dense层之前的，所以仅需要这个模型的前几层。

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
img (InputLayer)             [(None, 1280, 1024, 1)]   0         
_________________________________________________________________
conv2d (Conv2D)              (None, 639, 511, 2)       20        
_________________________________________________________________
batch_normalization (BatchNo (None, 639, 511, 2)       8         
_________________________________________________________________
activation (Activation)      (None, 639, 511, 2)       0         
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 319, 255, 2)       0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 159, 127, 4)       76        
_________________________________________________________________
batch_normalization_1 (Batch (None, 159, 127, 4)       16        
_________________________________________________________________
activation_1 (Activation)    (None, 159, 127, 4)       0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 79, 63, 4)         0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 77, 61, 4)         148       
_________________________________________________________________
batch_normalization_2 (Batch (None, 77, 61, 4)         16        
_________________________________________________________________
activation_2 (Activation)    (None, 77, 61, 4)         0         
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 38, 30, 4)         0         
_________________________________________________________________
global_average_pooling2d (Gl (None, 4)                 0         
_________________________________________________________________
dense (Dense)                (None, 10)                50        
_________________________________________________________________
dropout (Dropout)            (None, 10)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 10)                110       
_________________________________________________________________
dropout_1 (Dropout)          (None, 10)                0         
_________________________________________________________________
dense_2 (Dense)              (None, 1)                 11        
=================================================================
Total params: 455
Trainable params: 435
Non-trainable params: 20
_________________________________________________________________

构建方法

使用keras构建一般模型代码如下：

inp1 = Input(shape=(1280,1024,1),name="img")#输入层
out = Dense(20)(inp1)#一个全连接层
out = Dropout(0.5)(out)#dropout层
out = Dense(1)(out)#输出层

model = Model(inputs = inp1,outputs = out)#构建模型

将PreTrain好的模型的一些层放在自己的模型中代码如下：（迁移学习）

inp1 = Input(shape=(1280,1024,1),name="img")#输入层

#加入预训练的模型的卷积层，
conv = load_model('pre_train.h5')#此模型结构如图1
conv.trainable = False#设置让预训练好的模型的参数不更新
for i,layer in enumerate(conv.layers[1:-5]):#将conv模型的第一层到倒数第5层，也就是dense的前一层加入模型。
    if(i==0):
        out = layer(inp1)
    else:
        out = layer(out)
#后面接自己的网络层。
out = Dense(20)(out)#一个全连接层，将inp1改为out
out = Dropout(0.5)(out)#dropout层
out = Dense(1)(out)#输出层

model = Model(inputs = inp1,outputs = out)#构建模型

总结

这是最近用到的一个方法，学习记录一下，还有一种利用官方或者学术大佬公开模型参数进行迁移学习的，放在下一篇文章中记录。

PS：需要知道模型结构每层是什么网络层，方便后面按照层数放入自己的模型。