一、保存模型的全部配置信息

使用model.save()函数搭配tf.keras.models.load_model()对模型的架构，权重以及配置进行保存与恢复。

模型的保存代码如下：

import tensorflow as tf
import os

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['acc']
)

# 模型的训练
model.fit(train_images, train_labels, epochs=3)

# 模型的评测
print(model.evaluate(test_images, test_labels, verbose=0))

# 模型的保存
model.save(r'model_data/model.h5')

模型的恢复代码如下：

import tensorflow as tf
import os

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的恢复
model = tf.keras.models.load_model(r'model_data/model.h5')

# 模型的评测
print(model.evaluate(test_images, test_labels, verbose=0))

二、仅仅保存模型的架构

使用model.to_json()搭配tf.keras.models.model_from_json()对模型的架构进行保存与恢复

模型的保存代码如下：

import tensorflow as tf
import os
import json

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['acc']
)

# 模型的训练
model.fit(train_images, train_labels, epochs=3)

# 模型的评测
print(model.evaluate(test_images, test_labels, verbose=0))

# 模型的保存
# 生成json文件
model_json = model.to_json()
# 写入json文件
with open(r'model_json.json', 'w') as f:
    json.dump(model_json, f)
    print('模型的架构json文件保存完成！')

模型的恢复代码如下：

import tensorflow as tf
import os
import json

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的恢复
# 读取json文件
with open(r'model_json.json', 'r') as f:
    model_json = json.load(f)
# 模型的加载
model = tf.keras.models.model_from_json(model_json)
model.summary()

三、仅仅保存模型的权重

使用model.save_weights()函数搭配model.load_weights()函数对模型进行权重的保存与加载

模型权重的保存代码如下：

import tensorflow as tf
import os

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['acc']
)

# 模型的训练
model.fit(train_images, train_labels, epochs=3)

# 模型的评测
print(model.evaluate(test_images, test_labels, verbose=0))

# 模型的保存
model.save_weights(r'model_data/save_weights.h5')

模型权重的恢复代码如下：

import tensorflow as tf
import os

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['acc']
)

# 模型加载前的的评测
print('权重加载前的模型准确率')
print(model.evaluate(test_images, test_labels, verbose=0))
# 模型权重的加载
model.load_weights(r'model_data/save_weights.h5')
# 模型加载后的的评测
print('权重加载后的模型准确率')
print(model.evaluate(test_images, test_labels, verbose=0))

模型权重加载前后对比效果图

对比图

四、使用回调函数对模型进行保存

使用tf.keras.callbacks.ModelCheckpoint()回调函数对模型进行保存

模型权重的保存代码如下：

import tensorflow as tf
import os

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['acc']
)

# 设置回调函数
cp_callback = tf.keras.callbacks.ModelCheckpoint(r'checkpoint_data/logs',
                                                 save_weights_only=True)

# 模型的训练
model.fit(train_images, train_labels, epochs=3, callbacks=[cp_callback])

# 模型的评测
print(model.evaluate(test_images, test_labels, verbose=0))

模型的恢复代码如下：

import tensorflow as tf
import os

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['acc']
)

# 模型加载前的准确率
print(model.evaluate(test_images, test_labels, verbose=0))

# 模型的恢复
model.load_weights(r'checkpoint_data/logs')

# 模型加载后的准确率
print(model.evaluate(test_images, test_labels, verbose=0))

模型权重的恢复前后的对比图

对比图片

五、对于自定义训练的模型进行保存

使用tf.train.Checkpoint()函数的.save()方法与.restore()方法对模型进行保存与恢复

模型的保存代码如下：

import tensorflow as tf
import os
import tqdm

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()
dataset = tf.data.Dataset.from_tensor_slices((train_images, train_labels))
dataset = dataset.batch(60000)
test_dataset = tf.data.Dataset.from_tensor_slices((test_images, test_labels)).batch(60000)

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
optimizer = tf.keras.optimizers.Adam()
loss_func = tf.keras.losses.SparseCategoricalCrossentropy()

train_loss_mean = tf.keras.metrics.Mean('train_loss')
train_accuracy = tf.keras.metrics.Accuracy('train_accuracy')

test_loss_mean = tf.keras.metrics.Mean('test_loss')
test_accuracy = tf.keras.metrics.Accuracy('test_accuracy')

# 定义模型保存的函数
checkpoint = tf.train.Checkpoint(model=model)


# 定义单步的训练
def step_train(mol, images, labels):
    with tf.GradientTape() as t:
        pre = mol(images)
        loss_step = loss_func(labels, pre)
    grads = t.gradient(loss_step, mol.trainable_variables)
    optimizer.apply_gradients(zip(grads, mol.trainable_variables))
    train_loss_mean(loss_step)
    train_accuracy(labels, tf.argmax(pre, axis=-1))


def step_test(mol, imags, labels):
    pre = mol(imags, training=False)
    loss_step = loss_func(labels, pre)
    test_loss_mean(loss_step)
    test_accuracy(labels, tf.argmax(pre, axis=-1))


# 定义训练函数
def train():
    for i in range(300):
        tqdm_train = tqdm.tqdm(iter(dataset), total=len(dataset))
        for img, lab in tqdm_train:
            step_train(model, img, lab)
            tqdm_train.set_description_str('Epoch : {:3}'.format(i))
            tqdm_train.set_postfix_str(
                'train loss is {:.14f} train accuracy is {:.14f}'.format(train_loss_mean.result(),
                                                                         train_accuracy.result()))
        tqdm_test = tqdm.tqdm(iter(test_dataset), total=len(test_dataset))
        for ima, lbl in tqdm_test:
            step_test(model, ima, lbl)
            tqdm_test.set_description_str('Epoch : {:3}'.format(i))
            tqdm_test.set_postfix_str(
                'test loss is {:.14f} test accuracy is {:.14f}'.format(test_loss_mean.result(), test_accuracy.result()))
        if i % 50 == 0:
            checkpoint.save(file_prefix=r'save_check/logs')
        train_loss_mean.reset_states()
        train_accuracy.reset_states()
        test_loss_mean.reset_states()
        test_accuracy.reset_states()
        tqdm_train.close()
        tqdm_test.close()


if __name__ == '__main__':
    train()

模型的恢复代码如下：

import tensorflow as tf
import os
import tqdm

# 环境变量的配置
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

# 数据的加载
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.fashion_mnist.load_data()
test_dataset = tf.data.Dataset.from_tensor_slices((test_images, test_labels)).batch(60000)

# 模型的构建
model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.summary()

# 模型的相关配置
test_accuracy = tf.keras.metrics.Accuracy('test_accuracy')

# 定义模型保存的函数
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.restore(tf.train.latest_checkpoint(r'save_check'))


def step_test(mol, imags, labels):
    pre = mol(imags, training=False)
    test_accuracy(labels, tf.argmax(pre, axis=-1))


tqdm_test = tqdm.tqdm(iter(test_dataset), total=len(test_dataset))

for img, lable in tqdm_test:
    step_test(model, img, lable)
    tqdm_test.set_postfix_str(test_accuracy.result())
# tensorflow # 深度学习 # keras
Logo
华为云开发者联盟

为开发者提供学习成长、分享交流、生态实践、资源工具等服务，帮助开发者快速成长。

更多推荐

cover

如何基于香橙派AIpro对视频/图像数据进行预处理？

avatar 华为云开发者联盟
cover

HAS2024：华为云以系统性创新加速千行万业智能化升级

avatar 华为云开发者联盟
cover

数仓的两种轻量级数据交换格式：json与jsonb

avatar 华为云开发者联盟