Tensorflow-gpu保存keras模型为Savedmodel格式并载入进行推理

一、保存模型

mynetwork为keras模型对象，通过mynetwork = tf.keras.models.Sequential(...)类似方法建立；
tf.saved_model.save(mynetwork, 'saved_model')##保存模型，说明文件夹即可

二、载入保存的模型并进行推理

import tensorflow as tf
import numpy as np

gpus = tf.config.experimental.list_physical_devices('GPU')##获取可用GPU
for gpu in (gpus):
  tf.config.experimental.set_memory_growth(gpu, True)##设置显存使用方式

(x_train, y_train),(x_test, y_test) = tf.keras.datasets.mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0 ##数据预处理归一化

x_train = x_train[..., tf.newaxis] ##增加一个channle维数
x_test = x_test[..., tf.newaxis]

x_train = x_train.astype(np.float32)##原始数据类型为float64，因为调用了cudnn加速，需要转换为float32
x_test = x_test.astype(np.float32)


train_set = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(10000).batch(32)
test_set = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(32)


infer = tf.saved_model.load("saved_model")##读取模型

train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')##准确率定义
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')


for x, y in train_set:
    pre_train = infer.call(x)
    acc_train = train_accuracy(y, pre_train)
for x, y in test_set:
    pre_test = infer.call(x)
    acc_test = test_accuracy(y, pre_test)

print(acc_train)##输出准确率张量
print(acc_test)

三、预测结果