首先在看本内容之前，推荐你去看tensorflow官网的迁移学习和微调，那个讲的更清楚，也更加详细，网址如下所示。但是官网里面的模型加载确实在网络上下载的，不能符合实际的需求，因为有些情况，我们想用自己预训练的模型加载并且进行微调，那么就是本内容的重点。https://tensorflow.google.cn/guide/keras/transfer_learning?hl=zh_cn

大致的内容步骤如下

1、保存自己预训练的模型

2、加载自己的模型

3、进行微调

那么首先第一步，保存自己与训练的模型

我们首先应该调用model.save_weights()这个函数，保存自己的模型。数据集自自己调整下。

import tensorflow as tf
import numpy as np
from tensorflow.keras.layers import Conv1D,BatchNormalization,Dense,Activation,GlobalAveragePooling1D
from tensorflow.keras.regularizers import l2


auto_train_x_a = np.load("U_process_train_x.npy")[0:30,:]
auto_train_y = np.load("auto_process_data_train_y.npy")[0:30,:]

def auto():
    x_input = tf.keras.Input(shape=(128,))
    x1 = tf.keras.layers.Dense(128,kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x_input)
    x2 = tf.keras.layers.Dense(32, kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x1)
    x3 = tf.keras.layers.Dense(128, kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x2)

    model = tf.keras.Model(inputs=x_input,output=x3)

    return model


base_model = auto()
model.compile(''')
model.fit(''')
checkpoint_path = "auto/data_a/cp.ckpt"
base_model.save_weights(checkpoint_path)

那么首先第二步，加载自己训练的模型

import tensorflow as tf
import numpy as np
from tensorflow.keras.layers import Conv1D,BatchNormalization,Dense,Activation,GlobalAveragePooling1D
from tensorflow.keras.regularizers import l2


auto_train_x_a = np.load("U_process_train_x.npy")[0:30,:]
auto_train_y = np.load("auto_process_data_train_y.npy")[0:30,:]

def auto():
    x_input = tf.keras.Input(shape=(128,))
    x1 = tf.keras.layers.Dense(128,kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x_input)
    x2 = tf.keras.layers.Dense(32, kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x1)
    x3 = tf.keras.layers.Dense(128, kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x2)

    model = tf.keras.Model(inputs=x_input,output=x3)

    return model


base_model = auto()

checkpoint_path = "auto/data_a/cp.ckpt"
base_model.load_weights(checkpoint_path)

那么首先第三步，加载自己的模型后我们再添加新的层

import tensorflow as tf
import numpy as np
from tensorflow.keras.layers import Conv1D,BatchNormalization,Dense,Activation,GlobalAveragePooling1D
from tensorflow.keras.regularizers import l2


auto_train_x_a = np.load("U_process_train_x.npy")[0:30,:]
auto_train_y = np.load("auto_process_data_train_y.npy")[0:30,:]

def auto():
    x_input = tf.keras.Input(shape=(128,))
    x1 = tf.keras.layers.Dense(128,kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x_input)
    x2 = tf.keras.layers.Dense(32, kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x1)
    x3 = tf.keras.layers.Dense(128, kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x2)

    model = tf.keras.Model(inputs=x_input,output=x3)

    return model


base_model = auto()

checkpoint_path = "auto/data_a/cp.ckpt"
base_model.load_weights(checkpoint_path)
base_model.trainable = False  #(False是关闭梯度，不参与训练，Ture打开梯度进行训练)
x_input = tf.keras.Input(shape=(128,))
x = base_model(x_input)
x = tf.keras.layers.Dense(128,kernel_regularizer=tf.keras.regularizers.l2(1e-4))(x)
model2 = tf.keras.Model(inputs=x_input,output=x)

model2.compile()
model2.fit()