模型的保存与加载

save/load weights 只保存权值 需要源代码 轻量级save/load entire model 所有的状态保存saved_model 通用模型 所有的信息 不需要源代码 可以使用C++解析

方法一：save/load weights 只保存权值

# 保存权值 model.save_weights("./checkpoints/my_checkpoint") # 加载权值 model = create_model() # 重新创建装载模型 model.load_weights("./checkpoints/my_checkpoint") loss, acc = model.evaluate(test_images, test_labels) print("重新加载模型, 正确率: {:5.2f}%".format(100*acc))

实际使用状态

model.save_weights("./checkpoints/weights.ckpt") print("保存模型成功！") # 删除模型 del model # 重新创建模型 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)]) model.compile(optimizer=optimizers.Adam(lr=1e-3), loss=tf.losses.CategoricalCrossentropy(from_logits=True), metrics=["accuracy"]) # 加载模型 model.load_weights("./checkpoints/weights.ckpt") # 模型验证 loss, acc = model.evaluate(ds_val) print("重新加载模型, 正确率: {:5.2f}%".format(100*acc))

方法二：save model 保存模型与权值

不需要创建网络

model.save("model.h5") print("保存模型与权值成功！") del model print(“加载模型与权值”) model = tf.keras.models.load_model("model.h5") model.evaluate(ds_val)

方法三：saved_model通用模型(未测试过)

tf.saved_model.save(m, "/tmp/saved_model/") imported = tf.saved_model.load(path) f = imported.signatures["serving_default"] print(f(x = tf.ones([1, 28, 28, 3])))

方法一 完整代码

import tensorflow as tf from tensorflow.keras import Sequential, layers, optimizers import numpy as np # 获取物理GPU的个数 gpus = tf.config.experimental.list_physical_devices("GPU") for gpu in gpus: # 设置内存增长方式 自增长 tf.config.experimental.set_memory_growth(gpu, True) batch_size = 128 # 导入数据集 (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data() print(x_train.shape, y_train.shape, x_test.shape, y_test.shape) # 数据的预处理 def pre_process(x, y): x = tf.cast(x, tf.float32) / 255.0 x = tf.reshape(x, (28 * 28,)) y = tf.cast(y, tf.int32) y = tf.one_hot(y, depth=10) return x, y # 数据映射 打乱 分批 重复处理 db_train = tf.data.Dataset.from_tensor_slices((x_train, y_train)) db_train = db_train.map(pre_process).shuffle(60000).batch(batch_size).repeat() db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test)) db_test = db_test.map(pre_process).batch(batch_size) db_sample = iter(db_train) sample = next(db_sample) print("X:", sample[0].shape, "Y:", sample[1].shape) print(np.max(sample[0]), np.min(sample[0])) # 创建网络 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)]) model.build(input_shape=(None, 28 * 28)) model.summary() # 装载 设置 损失函数、优化器、评估指标 model.compile(optimizer=optimizers.Adam(lr=1e-3), loss=tf.losses.CategoricalCrossentropy(from_logits=True), metrics=["accuracy"]) # 循环训练 model.fit(db_train, # 指定数据集 epochs=5, # 循环次数 validation_data=db_test, # 测试数据集 validation_freq=2, # 验证频次 steps_per_epoch=x_train.shape[0] // batch_size) # 多余数据 # 评估测试 model.evaluate(db_test) model.save_weights("./checkpoints/weights.ckpt") print("保存模型成功！") # 删除模型 del model # 重新创建模型 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)]) model.compile(optimizer=optimizers.Adam(lr=1e-3), loss=tf.losses.CategoricalCrossentropy(from_logits=True), metrics=["accuracy"]) # 加载模型 model.load_weights("./checkpoints/weights.ckpt") # 模型验证 loss, acc = model.evaluate(db_test) print("重新加载模型, 正确率: {:5.2f}%".format(100 * acc))

方法二 完整代码

import tensorflow as tf from tensorflow.keras import Sequential, layers, optimizers import numpy as np # 获取物理GPU的个数 gpus = tf.config.experimental.list_physical_devices("GPU") for gpu in gpus: # 设置内存增长方式 自增长 tf.config.experimental.set_memory_growth(gpu, True) batch_size = 128 # 导入数据集 (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data() print(x_train.shape, y_train.shape, x_test.shape, y_test.shape) # 数据的预处理 def pre_process(x, y): x = tf.cast(x, tf.float32) / 255.0 x = tf.reshape(x, (28 * 28,)) y = tf.cast(y, tf.int32) y = tf.one_hot(y, depth=10) return x, y # 数据映射 打乱 分批 重复处理 db_train = tf.data.Dataset.from_tensor_slices((x_train, y_train)) db_train = db_train.map(pre_process).shuffle(60000).batch(batch_size).repeat() db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test)) db_test = db_test.map(pre_process).batch(batch_size) db_sample = iter(db_train) sample = next(db_sample) print("X:", sample[0].shape, "Y:", sample[1].shape) print(np.max(sample[0]), np.min(sample[0])) # 创建网络 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)]) model.build(input_shape=(None, 28 * 28)) model.summary() # 装载 设置 损失函数、优化器、评估指标 model.compile(optimizer=optimizers.Adam(lr=1e-3), loss=tf.losses.CategoricalCrossentropy(from_logits=True), metrics=["accuracy"]) # 循环训练 model.fit(db_train, # 指定数据集 epochs=3, # 循环次数 validation_data=db_test, # 测试数据集 validation_freq=2, # 验证频次 steps_per_epoch=x_train.shape[0] // batch_size) # 多余数据 # 评估测试 model.evaluate(db_test) save_path = "./model/model.h5" model.save(save_path) print("保存模型成功！") # 删除模型 del model # 加载模型 model = tf.keras.models.load_model(save_path) # 模型验证 loss, acc = model.evaluate(db_test) print("重新加载模型, 正确率: {:5.2f}%".format(100 * acc))