6.2 模型保存 --- 加载和保存模型结构&权重

一、只保存/加载模型的结构

保存模型的结构，而非其权重或训练配置项:
json_string = model.to_json()
 

model.save('my_model.h5')
my_model_json = model.to_json()
with open('my_json_model.json', 'w') as f:
    f.write(my_model_json)

from keras.models import model_from_json
model = model_from_json(json_string)

with open('my_json_model.json') as f:
    my_json_model = f.read()

from keras.models import model_from_json 
model = model_from_json(my_json_model)

model.summary()

# 只有架构，没有被编译，所以需要编译
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])


model.evaluate(test_image, test_label)

二、只保存/加载模型的权重

只保存模型的权重:
model.save_weights('my_model_weights.h5')

只加载模型的权重 :
model.load_weights('my_model_weights.h5')
model去加载权重，是使用已经定义好的model去加载权重

model.load_weights('my_model_weights.h5')
model.evaluate(test_image, test_label)

加载一部分权重

from keras import layers

model_new = keras.Sequential()
model_new.add(layers.Conv2D(64, (3, 3), activation='relu', input_shape=(28, 28, 1), name='conv_1'))
model_new.add(layers.Conv2D(64, (3, 3), activation='relu', name='conv_2'))
model_new.add(layers.MaxPooling2D(pool_size=(2, 2)))
model_new.add(layers.Flatten())
model_new.add(layers.Dense(256, activation='relu', name='dense_1_'))
model_new.add(layers.Dropout(0.5))
model_new.add(layers.Dense(10, activation='softmax', name='dense_2_'))

model.add(layers.Conv2D(64, (3, 3), activation='relu', input_shape=(28, 28, 1), name='conv_1'))
model.add(layers.Conv2D(64, (3, 3), activation='relu', name='conv_2'))
model.add(layers.MaxPooling2D(pool_size=(2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(256, activation='relu', name='dense_1'))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(10, activation='softmax', name='dense_2'))

model_new.summary()

model_new.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model_new.load_weights('my_model_weights.h5', by_name=True)

model_new.evaluate(test_image, test_label)

这样的话卷积层的权重加载进去了，但是Dense层是随机初始化的。