深度学习：TensorFlow2构建、保存、加载神经网络模型【经典流程】

一、network.save_weights、network.load_weights

保存模型的参数，加载已保存的参数的network的结构必须和之前的network的所有结构一模一样

import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'  # 放在 import tensorflow as tf 之前才有效

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers, optimizers, datasets

# 一、获取数据集
(X_train, Y_train), (X_test, Y_test) = datasets.mnist.load_data()
print('X_train.shpae = {0}，Y_train.shpae = {1}------------type(X_train) = {2}，type(Y_train) = {3}'.format(X_train.shape, Y_train.shape, type(X_train), type(Y_train)))


# 二、数据处理
# 预处理函数：将numpy数据转为tensor
def preprocess(x, y):
    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


# 2.1 处理训练集
# print('X_train.shpae = {0}，Y_train.shpae = {1}------------type(X_train) = {2}，type(Y_train) = {3}'.format(X_train.shape, Y_train.shape, type(X_train), type(Y_train)))
dataset_train = tf.data.Dataset.from_tensor_slices((X_train, Y_train))  # 此步骤自动将numpy类型的数据转为tensor
dataset_train = dataset_train.map(preprocess)  # 调用map()函数批量修改每一个元素数据的数据类型
dataset_train = dataset_train.shuffle(len(X_train))  # 打散dataset_train中的样本顺序，防止图片的原始顺序对神经网络性能的干扰
print('dataset_train = {0}，type(dataset_train) = {1}'.format(dataset_train, type(dataset_train)))
batch_size_train = 20000  # 每个batch里的样本数量设置100-200之间合适。
dataset_batch_train = dataset_train.batch(batch_size_train)  # 将dataset_batch_train中每sample_num_of_each_batch_train张图片分为一个batch，读取一个batch相当于一次性并行读取sample_num_of_each_batch_train张图片
print('dataset_batch_train = {0}，type(dataset_batch_train) = {1}'.format(dataset_batch_train, type(dataset_batch_train)))
# 2.2 处理测试集
dataset_test = tf.data.Dataset.from_tensor_slices((X_test, Y_test))  # 此步骤自动将numpy类型的数据转为tensor
dataset_test = dataset_test.map(preprocess)  # 调用map()函数批量修改每一个元素数据的数据类型
dataset_test = dataset_test.shuffle(len(X_test))  # 打散样本顺序，防止图片的原始顺序对神经网络性能的干扰
batch_size_test = 5000  # 每个batch里的样本数量设置100-200之间合适。
dataset_batch_test = dataset_test.batch(batch_size_test)  # 将dataset_test中每sample_num_of_each_batch_test张图片分为一个batch，读取一个batch相当于一次性并行读取sample_num_of_each_batch_test张图片

# 三、构建神经网络结构：Dense 表示全连接神经网络，激活函数用 relu
network = keras.Sequential([
    layers.Dense(500, activation=tf.nn.relu),  # 降维：784-->500
    layers.Dense(300, activation=tf.nn.relu),  # 降维：500-->300
    layers.Dense(100, activation=tf.nn.relu),  # 降维：300-->100
    layers.Dense(10)])  # 降维：100-->10，最后一层一般不需要在此处指定激活函数，在计算Loss的时候会自动运用激活函数
network.build(input_shape=[None, 784])  # 28*28=784，None表示样本数量，是不确定的值。
network.summary()  # 打印神经网络model的简要信息

# 四、设置神经网络各个参数
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )

# 五、给神经网络喂数据，训练神经网络模型参数
print('\n++++++++++++++++++++++++++++++++++++++++++++Training 阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
network.fit(dataset_batch_train, epochs=5, validation_data=dataset_batch_test, validation_freq=2)  # validation_freq参数表示每多少个epoch做一次验证/validation
print('++++++++++++++++++++++++++++++++++++++++++++Training 阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

# 六、模型评估 test/evluation
print('\n++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
network.evaluate(dataset_batch_test)
print('++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

network.save_weights('weights.ckpt')
print('\n================saved weights================')
del network
print('================del network================')

# 七、创建一个和所加载参数的原始network一模一样的network
print('================创建一个和所加载参数的原始network一模一样的network================')
network = keras.Sequential([
    layers.Dense(500, activation=tf.nn.relu),  # 降维：784-->500
    layers.Dense(300, activation=tf.nn.relu),  # 降维：500-->300
    layers.Dense(100, activation=tf.nn.relu),  # 降维：300-->100
    layers.Dense(10)])  # 降维：100-->10，最后一层一般不需要在此处指定激活函数，在计算Loss的时候会自动运用激活函数
network.build(input_shape=[None, 784])  # 28*28=784，None表示样本数量，是不确定的值。
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )
network.load_weights('weights.ckpt')
print('================loaded weights================')

# 八、模型评估 test/evluation
print('\n++++++++++++++++++++++++++++++++++++++++++++加载weights后--->Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
network.evaluate(dataset_batch_test)
print('++++++++++++++++++++++++++++++++++++++++++++加载weights后--->Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

# 九、模型上线应用
sample = next(iter(dataset_batch_test))  # 从 dataset_batch_test 中取一个batch数据做模拟
x = sample[0]
y = sample[1]  # one-hot
pred = network.predict(x)  # [b, 10]
y = tf.argmax(y, axis=1)  # convert back to number
pred = tf.argmax(pred, axis=1)
print('\n++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
print(pred)
print(y)
print('++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

打印结果：

X_train.shpae = (60000, 28, 28)，Y_train.shpae = (60000,)------------type(X_train) = <class 'numpy.ndarray'>，type(Y_train) = <class 'numpy.ndarray'>
dataset_train = <ShuffleDataset shapes: ((784,), (10,)), types: (tf.float32, tf.float32)>，type(dataset_train) = <class 'tensorflow.python.data.ops.dataset_ops.ShuffleDataset'>
dataset_batch_train = <BatchDataset shapes: ((None, 784), (None, 10)), types: (tf.float32, tf.float32)>，type(dataset_batch_train) = <class 'tensorflow.python.data.ops.dataset_ops.BatchDataset'>
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 500)               392500    
_________________________________________________________________
dense_1 (Dense)              (None, 300)               150300    
_________________________________________________________________
dense_2 (Dense)              (None, 100)               30100     
_________________________________________________________________
dense_3 (Dense)              (None, 10)                1010      
=================================================================
Total params: 573,910
Trainable params: 573,910
Non-trainable params: 0
_________________________________________________________________

++++++++++++++++++++++++++++++++++++++++++++Training 阶段：开始++++++++++++++++++++++++++++++++++++++++++++
Epoch 1/5
3/3 [==============================] - 2s 113ms/step - loss: 2.7174 - accuracy: 0.1086
Epoch 2/5
3/3 [==============================] - 3s 492ms/step - loss: 2.6596 - accuracy: 0.1666 - val_loss: 1.6333 - val_accuracy: 0.4709
Epoch 3/5
3/3 [==============================] - 2s 115ms/step - loss: 1.5516 - accuracy: 0.4968
Epoch 4/5
3/3 [==============================] - 2s 255ms/step - loss: 1.0690 - accuracy: 0.6475 - val_loss: 0.7587 - val_accuracy: 0.7859
Epoch 5/5
3/3 [==============================] - 2s 115ms/step - loss: 0.7137 - accuracy: 0.7955
++++++++++++++++++++++++++++++++++++++++++++Training 阶段：结束++++++++++++++++++++++++++++++++++++++++++++

++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++
2/2 [==============================] - 0s 22ms/step - loss: 0.5240 - accuracy: 0.8493
++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++

================saved weights================
================del network================
================创建一个和所加载参数的原始network一模一样的network================
================loaded weights================

++++++++++++++++++++++++++++++++++++++++++++加载weights后--->Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++
2/2 [==============================] - 0s 22ms/step - loss: 0.5223 - accuracy: 0.8486
++++++++++++++++++++++++++++++++++++++++++++加载weights后--->Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++

++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：开始++++++++++++++++++++++++++++++++++++++++++++
tf.Tensor([6 3 7 ... 5 1 0], shape=(5000,), dtype=int64)
tf.Tensor([6 3 7 ... 3 1 0], shape=(5000,), dtype=int64)
++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：结束++++++++++++++++++++++++++++++++++++++++++++

Process finished with exit code 0

二、network.save()、network.load()

保存整个模型，加载后再根据network的通常做法进行操作。

import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'  # 放在 import tensorflow as tf 之前才有效

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers, optimizers, datasets

# 一、获取数据集
(X_train, Y_train), (X_test, Y_test) = datasets.mnist.load_data()
print('X_train.shpae = {0}，Y_train.shpae = {1}------------type(X_train) = {2}，type(Y_train) = {3}'.format(X_train.shape, Y_train.shape, type(X_train), type(Y_train)))


# 二、数据处理
# 预处理函数：将numpy数据转为tensor
def preprocess(x, y):
    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


# 2.1 处理训练集
# print('X_train.shpae = {0}，Y_train.shpae = {1}------------type(X_train) = {2}，type(Y_train) = {3}'.format(X_train.shape, Y_train.shape, type(X_train), type(Y_train)))
dataset_train = tf.data.Dataset.from_tensor_slices((X_train, Y_train))  # 此步骤自动将numpy类型的数据转为tensor
dataset_train = dataset_train.map(preprocess)  # 调用map()函数批量修改每一个元素数据的数据类型
dataset_train = dataset_train.shuffle(len(X_train))  # 打散dataset_train中的样本顺序，防止图片的原始顺序对神经网络性能的干扰
print('dataset_train = {0}，type(dataset_train) = {1}'.format(dataset_train, type(dataset_train)))
batch_size_train = 20000  # 每个batch里的样本数量设置100-200之间合适。
dataset_batch_train = dataset_train.batch(batch_size_train)  # 将dataset_batch_train中每sample_num_of_each_batch_train张图片分为一个batch，读取一个batch相当于一次性并行读取sample_num_of_each_batch_train张图片
print('dataset_batch_train = {0}，type(dataset_batch_train) = {1}'.format(dataset_batch_train, type(dataset_batch_train)))
# 2.2 处理测试集
dataset_test = tf.data.Dataset.from_tensor_slices((X_test, Y_test))  # 此步骤自动将numpy类型的数据转为tensor
dataset_test = dataset_test.map(preprocess)  # 调用map()函数批量修改每一个元素数据的数据类型
dataset_test = dataset_test.shuffle(len(X_test))  # 打散样本顺序，防止图片的原始顺序对神经网络性能的干扰
batch_size_test = 5000  # 每个batch里的样本数量设置100-200之间合适。
dataset_batch_test = dataset_test.batch(batch_size_test)  # 将dataset_test中每sample_num_of_each_batch_test张图片分为一个batch，读取一个batch相当于一次性并行读取sample_num_of_each_batch_test张图片

# 三、构建神经网络结构：Dense 表示全连接神经网络，激活函数用 relu
network = keras.Sequential([
    layers.Dense(500, activation=tf.nn.relu),  # 降维：784-->500
    layers.Dense(300, activation=tf.nn.relu),  # 降维：500-->300
    layers.Dense(100, activation=tf.nn.relu),  # 降维：300-->100
    layers.Dense(10)])  # 降维：100-->10，最后一层一般不需要在此处指定激活函数，在计算Loss的时候会自动运用激活函数
network.build(input_shape=[None, 784])  # 28*28=784，None表示样本数量，是不确定的值。
network.summary()  # 打印神经网络model的简要信息

# 四、设置神经网络各个参数
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )

# 五、给神经网络喂数据，训练神经网络模型参数
print('\n++++++++++++++++++++++++++++++++++++++++++++Training 阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
network.fit(dataset_batch_train, epochs=5, validation_data=dataset_batch_test, validation_freq=2)  # validation_freq参数表示每多少个epoch做一次验证/validation
print('++++++++++++++++++++++++++++++++++++++++++++Training 阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

# 六、模型评估 test/evluation
print('\n++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
network.evaluate(dataset_batch_test)
print('++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

network.save('model.h5')
print('\n================saved total model================')
del network
print('================del network================')

# 七、从磁盘加载保存的整体模型(包括所有参数、结构...)
print('================loaded model from file================')
network = tf.keras.models.load_model('model.h5', compile=False)
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )

# 八、模型评估 test/evluation
print('\n++++++++++++++++++++++++++++++++++++++++++++从磁盘加载整个model后--->Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
network.evaluate(dataset_batch_test)
print('++++++++++++++++++++++++++++++++++++++++++++从磁盘加载整个model后--->Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

# 九、模型上线应用
sample = next(iter(dataset_batch_test))  # 从 dataset_batch_test 中取一个batch数据做模拟
x = sample[0]
y = sample[1]  # one-hot
pred = network.predict(x)  # [b, 10]
y = tf.argmax(y, axis=1)  # convert back to number
pred = tf.argmax(pred, axis=1)
print('\n++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：开始++++++++++++++++++++++++++++++++++++++++++++')
print(pred)
print(y)
print('++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：结束++++++++++++++++++++++++++++++++++++++++++++')

打印结果：

X_train.shpae = (60000, 28, 28)，Y_train.shpae = (60000,)------------type(X_train) = <class 'numpy.ndarray'>，type(Y_train) = <class 'numpy.ndarray'>
dataset_train = <ShuffleDataset shapes: ((784,), (10,)), types: (tf.float32, tf.float32)>，type(dataset_train) = <class 'tensorflow.python.data.ops.dataset_ops.ShuffleDataset'>
dataset_batch_train = <BatchDataset shapes: ((None, 784), (None, 10)), types: (tf.float32, tf.float32)>，type(dataset_batch_train) = <class 'tensorflow.python.data.ops.dataset_ops.BatchDataset'>
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 500)               392500    
_________________________________________________________________
dense_1 (Dense)              (None, 300)               150300    
_________________________________________________________________
dense_2 (Dense)              (None, 100)               30100     
_________________________________________________________________
dense_3 (Dense)              (None, 10)                1010      
=================================================================
Total params: 573,910
Trainable params: 573,910
Non-trainable params: 0
_________________________________________________________________

++++++++++++++++++++++++++++++++++++++++++++Training 阶段：开始++++++++++++++++++++++++++++++++++++++++++++
Epoch 1/5
3/3 [==============================] - 2s 119ms/step - loss: 2.4869 - accuracy: 0.2464
Epoch 2/5
3/3 [==============================] - 2s 514ms/step - loss: 3.5169 - accuracy: 0.3786 - val_loss: 1.5471 - val_accuracy: 0.5026
Epoch 3/5
3/3 [==============================] - 2s 116ms/step - loss: 1.4532 - accuracy: 0.5238
Epoch 4/5
3/3 [==============================] - 2s 273ms/step - loss: 0.9930 - accuracy: 0.6789 - val_loss: 0.6357 - val_accuracy: 0.8010
Epoch 5/5
3/3 [==============================] - 2s 112ms/step - loss: 0.6005 - accuracy: 0.8118
++++++++++++++++++++++++++++++++++++++++++++Training 阶段：结束++++++++++++++++++++++++++++++++++++++++++++

++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++
2/2 [==============================] - 0s 24ms/step - loss: 0.4489 - accuracy: 0.8735
++++++++++++++++++++++++++++++++++++++++++++Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++

================saved total model================
================del network================
================loaded model from file================

++++++++++++++++++++++++++++++++++++++++++++从磁盘加载整个model后--->Evluation 阶段：开始++++++++++++++++++++++++++++++++++++++++++++
2/2 [==============================] - 0s 21ms/step - loss: 0.4505 - accuracy: 0.8729
++++++++++++++++++++++++++++++++++++++++++++从磁盘加载整个model后--->Evluation 阶段：结束++++++++++++++++++++++++++++++++++++++++++++

++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：开始++++++++++++++++++++++++++++++++++++++++++++
tf.Tensor([9 0 9 ... 5 1 9], shape=(5000,), dtype=int64)
tf.Tensor([9 0 9 ... 5 1 9], shape=(5000,), dtype=int64)
++++++++++++++++++++++++++++++++++++++++++++加载weights后--->应用阶段：结束++++++++++++++++++++++++++++++++++++++++++++

Process finished with exit code 0

三、tf.saved_model.save()、tf.saved_model.load()、

保存为可以被其他语言(比如：C++)调用的格式