如何重用keras功能模型的层级

Posted 2023-02-16

【中文标题】如何重用keras功能模型的层级

【英文标题】：How to reuse the layers of keras functional model

【发布时间】：2021-12-20 14:44:25

【问题描述】：

import tensorflow as tf
import keras

def get_model():

    x1 = keras.layers.Dense(6, activation='relu',input_shape=(10,))
    x2 = keras.layers.Dense(3, activation='relu')(x1)
    output_ = keras.layers.Dense(10,acitvation='sigmoid')(x2)

    model = keras.model(inputs=[x1], outputs=[output_])

    return model

model = get_model()

model.compile(...)

chk_point = keras.callbacks.ModelCheckpoint(f'./best_model.h5', 
                            monitor='val_loss', save_best_only=True, mode='min')
model.fit(..., callbacks=[chk_point])


def new_model():
    old = '../best_model.h5' #using old model for training new model

（我正在尝试使用预训练的功能模型进行迁移学习）

现在我想获取 best_model 的所有层。如果可能的话，我想删除我的 best_model 的最后一层。我想冻结所有层 best_model 即可训练 = 假。并向该模型添加新层。

我正在尝试在输入和 best_model.h5 的输出相同（eg.input_shape=(100,) 和 output_shape=(100,))。然后我正在考虑冻结所有这些层并在向该模型添加新层之后删除该模型的最后一层。 然后像往常一样在 X 和 y 上训练该模型

【参考方案1】：

一种方法是定义新模型，然后从 旧 模型中复制层权重（最后一层除外）并将 trainable 设置为 False。例如，假设您要删除最后一层并添加两个密集层（这只是一个示例）。请注意，当前模型的输入和输出大小为 (10,)。 另请注意，功能 API 中的第一层是输入层。 代码如下：

import tensorflow as tf
from tensorflow import keras
from sklearn.model_selection import train_test_split
import numpy as np


def get_model():

    inputs = keras.Input(shape=(10, ))
    x1 = keras.layers.Dense(6, activation='relu')(inputs)
    x2 = keras.layers.Dense(3, activation='relu')(x1)
    output_ = keras.layers.Dense(10,activation='sigmoid')(x2)

    model = keras.Model(inputs=inputs, outputs=[output_])

    return model


def get_new_model():

    inputs = keras.Input(shape=(10, ))
    x1 = keras.layers.Dense(6, activation='relu')(inputs)
    x2 = keras.layers.Dense(3, activation='relu')(x1)

    # new layers
    x3 = keras.layers.Dense(15, activation='relu')(x2)
    output_ = keras.layers.Dense(10, activation='sigmoid')(x3)

    model = keras.Model(inputs=inputs, outputs=[output_])

    return model


model = get_model()
model.compile(optimizer='adam', loss='mse')
batch_size = 16
_ = model.call(inputs=tf.random.normal(shape=(batch_size, 10)))
model.summary()

# create x data using two normal distributions with different mean
# y data is unused in auto encoder
x0 = np.random.normal(loc=0.0, size=(100, 10))
x1 = np.random.normal(loc=0.3, size=(100, 10))
x = np.concatenate((x0, x1), axis=0)
# y is unused
y0 = np.zeros((100, 10))
y1 = np.ones((100, 10))
y = np.concatenate((y0, y1), axis=0)
# split train/validation data
x_train, x_val, y_train, y_val = train_test_split(x, y, train_size=0.7) 
print(x_train.shape)
print(y_train.shape)

chk_point = keras.callbacks.ModelCheckpoint(f'./best_model.h5', 
                            monitor='val_loss', save_best_only=True, mode='min')
history = model.fit(x=x_train, y=x_train, batch_size=batch_size, epochs=100, callbacks=[chk_point], validation_data=(x_val, x_val))

# reload old model
model_old = keras.models.load_model('./best_model.h5')
model_old.summary()

# get new model
model_new = get_new_model()
model_new.compile(optimizer='adam', loss='mse')
_ = model_new.call(inputs=tf.random.normal(shape=(batch_size, 10)))
model_new.summary()

# copy the two dense layer weights and set trainable to False
# skip the first layer which is an InputLayer
for count, (layer_old, layer_new) in enumerate(zip(model_old.layers[1:3], model_new.layers[1:3])):  
  layer_new.trainable = False
  layer_new.set_weights(layer_old.get_weights())
  model_new.layers[count+1] = layer_new

model_new.summary()