【Keras】保存模型的前几层，删除最后几层

1 前言

⭐️ 需求：在使用Keras的过程中，只想保留模型的前几层，删除最后一层，以便网络进行增量训练。
⭐️ 以sklearn中的鸢尾花数据集为例，建立一个多层感知机，以用来删除网络的最后一层
⭐️ 使用以下代码进行删除训练好神经网络的最后一层

model.pop()

2 代码

2.1 删除前代码

⭐️ 载入数据集

import tensorflow as tf
import numpy as np
from sklearn.datasets import load_iris
from tqdm import tqdm

data = load_iris()
iris_target = data.target
iris_data = np.float32(data.data)

⭐️ 建立模型并训练

from sklearn.model_selection import train_test_split
from keras.layers import Dense, Dropout
import tensorflow as tf
import keras
from keras.models import Sequential
from tensorflow.keras.optimizers import SGD, Adam


X_train, X_test, y_train, y_test = train_test_split(iris_data, iris_target, test_size = 0.25)

def build_mlp(max_features):
    model = Sequential()
    model.add(Dense(units = 30, input_dim = max_features, activation = "relu"))
    model.add(Dropout(0.2))
    model.add(Dense(units = 30, activation = "relu"))
    model.add(Dense(units = 3, activation = "softmax"))
    # sgd = SGD(lr = 0.002)

    model.compile(optimizer="adam", loss=tf.losses.sparse_categorical_crossentropy, metrics = ["accuracy"])
    return model
# 设置模型提前停止训练
callback = keras.callbacks.EarlyStopping(monitor='loss', patience=5)
model = build_mlp(4)
model.fit(X_train, y_train, epochs=10,  batch_size=128, verbose=1, callbacks = [callback])
score1 = model.evaluate(X_test, y_test)
print(score1)

2.2 删除前模型

⭐️ 查看模型结构代码

model.summary()

⭐️ 模型结构

Model: "sequential_4"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_12 (Dense)            (None, 30)                150       
                                                                 
 dropout_3 (Dropout)         (None, 30)                0         
                                                                 
 dense_13 (Dense)            (None, 30)                930       
                                                                 
 dense_14 (Dense)            (None, 3)                 93        
                                                                 
=================================================================
Total params: 1,173
Trainable params: 1,173
Non-trainable params: 0
_________________________________________________________________

⭐️ 模型预测代码及结果

model.predict([[5.1, 3.5, 1.4, 0.2]])
# 输出结果为：array([[0.15257263, 0.56145775, 0.2859696 ]], dtype=float32)

2.3 删除最后一层代码

⭐️ 删除代码

model.pop()

2.4 删除后结果

⭐️ 查看删除后模型结构

model.summary()

⭐️ 删除后模型结构

Model: "sequential_4"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_12 (Dense)            (None, 30)                150       
                                                                 
 dropout_3 (Dropout)         (None, 30)                0         
                                                                 
 dense_13 (Dense)            (None, 30)                930       
                                                                 
=================================================================
Total params: 1,080
Trainable params: 1,080
Non-trainable params: 0
_________________________________________________________________

⭐️ 删除最后一层后模型预测代码及结果

model.save("test_model.h5")
new_model = tf.keras.models.load_model("test_model.h5")
new_model.predict([[5.1, 3.5, 1.4, 0.2]])

⭐️ 输出结果

array([[0.6801046 , 0.        , 1.6479366 , 2.1548553 , 0.        ,
        0.        , 0.        , 0.        , 0.3491744 , 1.1407335 ,
        2.2483244 , 0.        , 0.71208394, 0.23256747, 0.        ,
        0.        , 1.1044086 , 1.492352  , 0.        , 0.        ,
        1.6636133 , 0.0603824 , 0.02222613, 0.        , 0.        ,
        1.625694  , 0.        , 0.        , 0.        , 3.2165537 ]],
      dtype=float32)

3 总结

⭐️ pop()函数是一层层删除训练后的模型，从后往前删除，就像栈的机制。

4 参考资料

📗 知乎连接：keras可以只保存模型前几层，而不是整个模型及权重值吗？