tensorflow2 保存模型

__author__ = 'LXY'
import numpy as np
import tensorflow as tf
from tensorflow.keras import datasets,layers,optimizers,Sequential,metrics
tf.random.set_seed(1)
#数据预处理
def preprocess(x,y):
    #转换x,y的数据类型为张量
    x = tf.cast(x,dtype=tf.float32)/255.
    x = tf.reshape(x,[28*28])
    y = tf.cast(y,dtype=tf.int32)
    y = tf.one_hot(y,depth=10)
    return x,y

def load_mnist_data():
    batchsz = 128
    (x_train, y_train), (x_test, y_test) = datasets.mnist.load_data()
    db_train = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    db_train = db_train.map(preprocess).shuffle(60000).batch(batch_size=batchsz)

    db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test))
    db_test = db_test.map(preprocess).batch(batch_size=batchsz)
    return db_train, db_test

def build_model():
    db_train, db_test = load_mnist_data()

    sample = next(iter(db_train))
    # 设计网络模型
    model = Sequential([
        layers.Dense(256, activation='relu'),
        layers.Dense(128, activation='relu'),
        layers.Dense(64, activation='relu'),
        layers.Dense(32, activation='relu'),
        layers.Dense(10)
    ])
    # 建立模型
    model.build(input_shape=[None, 28 * 28])
    #model.summary()  # 输出网络结构
    # 编译
    model.compile(
        optimizer=optimizers.Adam(lr=0.001),  # 能够直接重定义lr
        loss=tf.losses.CategoricalCrossentropy(from_logits=True),
        metrics=['accuracy', 'mse']  # 通过运行，训练数据后返回的history
    )

    history = model.fit(db_train, epochs=2, validation_data=db_train, validation_freq=2, verbose=0)
    acc = history.history['accuracy']
    mse = history.history['mse']
    # 从训练结果中拿想看的字段，比如输出每次训练
    print(history.history, '\nacc', acc, '\nmse', mse)
    # 输出是每次迭代的acc和mse

    # test
    index_test = model.evaluate(db_test, verbose=0)

    print("index_test", model.metrics_names, index_test)
    
    print("acc_test", index_test[0], "\nmse_test", index_test[1])

    sample = next(iter(db_test))

    x = sample[0]
    y = sample[1]
    predict = model.predict(x)

    y = tf.argmax(y, axis=1)
    predict = tf.argmax(predict, axis=1)
    result = tf.equal(y, predict)
    result = tf.cast(result, tf.int32)
    num_correct = tf.reduce_sum(result)
    print("正确率", float(num_correct / predict.shape[0]))  # 预测正确的概率
    return model

if __name__=='__main__':
    model = build_model()
    #保存模型
    model.save('mnist_model')
    #删除model
    del model

    #重新加载模型
    network = tf.keras.models.load_model('mnist_model',compile=True)
    train_data,test_data = load_mnist_data()

    result_test = network.evaluate(test_data)
    print('result_test',network.metrics_names, result_test)

 模型保存

模型保存 model.save("name")

model.save('mnist_model')

当训练好了，拿出来再用的话：

network = tf.keras.models.load_model('mnist_model',compile=True)

想要test数据，直接使用模型就可以了

输出结果：

{'loss': [0.28666120767593384, 0.10383852571249008], 'accuracy': [0.9153000116348267, 0.968833327293396], 'mse': [14.382641792297363, 20.249595642089844], 'val_loss': [0.07012375444173813], 'val_accuracy': [0.9796333312988281], 'val_mse': [19.037185668945312]} 
acc [0.9153000116348267, 0.968833327293396] 
mse [14.382641792297363, 20.249595642089844]
index_test ['loss', 'accuracy', 'mse'] [0.09124262630939484, 0.972100019454956, 19.534870147705078]
acc_test 0.09124262630939484 
mse_test 0.972100019454956
正确率 0.984375
#下面是使用模型evaluate的结果
79/79 [==============================] - 0s 4ms/step - loss: 0.0912 - accuracy: 0.9721 - mean_squared_error: 19.5349
result_test ['loss', 'accuracy', 'mean_squared_error'] [0.09124262630939484, 0.972100019454956, 19.534870147705078]

如果在main函数当中，要重新修改模型的优化函数，或者返回的性能指标的话，令compile=false,

if __name__=='__main__':
    model = build_model()
    #保存模型
    model.save('mnist_model')
    #删除model
    del model

    #重新加载模型
    network = tf.keras.models.load_model('mnist_model',compile=False)
#重新定义compile
    network.compile(
        optimizer = tf.keras.optimizers.Adam(lr=0.02),
        loss = tf.losses.CategoricalCrossentropy(from_logits=True),
        metrics=['accuracy', 'mse']  # 通过运行，训练数据后返回的history
    )
    train_data,test_data = load_mnist_data()

    result_test = network.evaluate(test_data)
    print('result_test',network.metrics_names, result_test)

输出结果

{'loss': [0.28666120767593384, 0.10383852571249008], 'accuracy': [0.9153000116348267, 0.968833327293396], 'mse': [14.382641792297363, 20.249595642089844], 'val_loss': [0.07012375444173813], 'val_accuracy': [0.9796333312988281], 'val_mse': [19.037185668945312]} 
acc [0.9153000116348267, 0.968833327293396] 
mse [14.382641792297363, 20.249595642089844]
index_test ['loss', 'accuracy', 'mse'] [0.09124262630939484, 0.972100019454956, 19.534870147705078]
acc_test 0.09124262630939484 
mse_test 0.972100019454956
正确率 0.984375
#compile修改后的输出结果
79/79 [==============================] - 0s 4ms/step - loss: 0.0912 - accuracy: 0.9721 - mse: 19.5349
result_test ['loss', 'accuracy', 'mse'] [0.09124262630939484, 0.972100019454956, 19.534870147705078]

 可以看到，输出结果两次都一毛一样，哈哈哈哈哈哈哈哈

因为新改的compile()函数并没有被执行到，

再加一句：

history = network.fit(train_data, epochs=2, validation_data=test_data, validation_freq=2, verbose=0)

main函数变成：

if __name__=='__main__':
    model = build_model()
    #保存模型
    model.save('mnist_model')
    #删除model
    del model

    #重新加载模型
    network = tf.keras.models.load_model('mnist_model',compile=False)
    network.compile(
        optimizer = tf.keras.optimizers.Adam(lr=0.02),
        loss = tf.losses.CategoricalCrossentropy(from_logits=True),
        metrics=['accuracy', 'mse']  # 通过运行，训练数据后返回的history
    )
    train_data,test_data = load_mnist_data()
    history = network.fit(train_data, epochs=2, validation_data=test_data, validation_freq=2, verbose=0)
    result_test = network.evaluate(test_data)
    print('result_test',network.metrics_names, result_test)

输出结果：

{'loss': [0.28666120767593384, 0.10383852571249008], 'accuracy': [0.9153000116348267, 0.968833327293396], 'mse': [14.382641792297363, 20.249595642089844], 'val_loss': [0.07012375444173813], 'val_accuracy': [0.9796333312988281], 'val_mse': [19.037185668945312]} 
acc [0.9153000116348267, 0.968833327293396] 
mse [14.382641792297363, 20.249595642089844]
index_test ['loss', 'accuracy', 'mse'] [0.09124262630939484, 0.972100019454956, 19.534870147705078]
acc_test 0.09124262630939484 
mse_test 0.972100019454956
正确率 0.984375
#变了
79/79 [==============================] - 0s 3ms/step - loss: 0.1924 - accuracy: 0.9463 - mse: 45.6422
result_test ['loss', 'accuracy', 'mse'] [0.19236354529857635, 0.9463000297546387, 45.64221954345703]

前一段数据没变，是因为输出的一直都是原来保存的模型的输出结果，后面新改的compile()要fit一下才起作用，否则就只是把训练好的模型用来测试了。

参数保存

如果再把模型的参数保存下，用的时候再提出来：

if __name__=='__main__':
    model = build_model()
    #保存模型
    model.save('mnist_model')
    model.save_weights('weights')
    #删除model
    del model

    #重新加载模型
    network = tf.keras.models.load_model('mnist_model',compile=False)
    network.compile(
        optimizer = tf.keras.optimizers.Adam(lr=0.02),
        loss = tf.losses.CategoricalCrossentropy(from_logits=True),
        metrics=['accuracy', 'mse']  # 通过运行，训练数据后返回的history
    )
    train_data,test_data = load_mnist_data()
    network.load_weights('weights')
    history = network.fit(train_data, epochs=2, validation_data=test_data, validation_freq=2, verbose=0)
    result_test = network.evaluate(test_data)
    print('result_test',network.metrics_names, result_test)

 输出结果：

#又变了的结果
79/79 [==============================] - 0s 4ms/step - loss: 0.0718 - accuracy: 0.9794 - mse: 28.5494
result_test ['loss', 'accuracy', 'mse'] [0.07176566123962402, 0.9793999791145325, 28.549360275268555]

这部分的loss比上面更低，更精准了，连mse都低了很多。