tf2 模型保存总结

tf2 模型保存总结

1. model.save保存的是所有信息，结果是单文件，最为简单。

实例：保 加

model_name = "./model_save/fassionMnist_save.h5"

model.save(model_name)

new_model = keras.models.load_model(model_name)

2. model.save_weights(weight_file)保存的是权重，结果是单文件。

weight_file="./model_save/weights.h5"

示例：保 创 编 加

model.save_weights(weight_file)

?

model = keras.Sequential()

model.add(keras.layers.Flatten(input_shape=(28,28)))

model.add(keras.layers.Dense(128,activation="relu"))

model.add(keras.layers.Dense(10, activation="softmax"))

model.summary()

?

model.compile(optimizer="adam",

loss="sparse_categorical_crossentropy",

metrics=["acc"])

?

model.load_weights(weight_file)

3. 检查点保存权重，结果多文件

示例：

ckpt_path="./ckpt/model_ckpt.ckpt"

ckpt_callback=keras.callbacks.ModelCheckpoint(

ckpt_path,save_weights_only=True)

history = model.fit(train_image,train_label,epochs=3,callbacks=[ckpt_callback])

?

model = keras.Sequential()

model.add(keras.layers.Flatten(input_shape=(28,28)))

model.add(keras.layers.Dense(128,activation="relu"))

model.add(keras.layers.Dense(10, activation="softmax"))

model.summary()

?

model.compile(optimizer="adam",

loss="sparse_categorical_crossentropy",

metrics=["acc"])

?

model.load_weights(ckpt_path)

?

4. 检查点保存全部模型，结果是文件夹

而且win下保存路径必须用 反斜杠，不能用正斜杠，可视为bug

model_ckpt_path=".\ckpt\model3.model"

ckpt_callback=keras.callbacks.ModelCheckpoint(

model_ckpt_path,save_weights_only=False)

model.evaluate(test_image,test_label,verbose=0)

history = model.fit(train_image,train_label,epochs=3,callbacks=[ckpt_callback])

model.evaluate(test_image,test_label,verbose=0)

?

new_model = keras.models.load_model(model_ckpt_path)

new_model.evaluate(test_image,test_label,verbose=0)

tf2 模型保存总结

原文：https://www.cnblogs.com/xiexiaokui/p/12268031.html