专家入门TensorFlow 2.0使用流程：数据处理、自定义模型、损失、指标、梯度下降 (tensorflow2.0官方教程翻译)

初学者入门教程中，使用tf.keras.Sequential模型，只是简单的堆叠模型。 本文是专家级入门，使用 Keras 模型子类 API 构建模型，会使用更底层一点的的函数接口，自定义模型、损失、评估指标和梯度下降控制等，流程清晰。

开始，请将TensorFlow库导入您的程序：

​x
1
from __future__ import absolute_import, division, print_function, unicode_literals
2
​
3
import tensorflow as tf  # 安装命令 `pip install tensorflow-gpu==2.0.0-alpha0`
4
​
5
from tensorflow.keras.layers import Dense, Flatten, Conv2D
6
from tensorflow.keras import Model

加载并准备MNIST数据集.。

xxxxxxxxxx
8
1
mnist = tf.keras.datasets.mnist
2
​
3
(x_train, y_train), (x_test, y_test) = mnist.load_data()
4
x_train, x_test = x_train / 255.0, x_test / 255.0
5
​
6
# 添加一个通道维度
7
x_train = x_train[..., tf.newaxis]
8
x_test = x_test[..., tf.newaxis]

使用tf.data批处理和随机打乱数据集：

xxxxxxxxxx
3
1
train_ds = tf.data.Dataset.from_tensor_slices(
2
    (x_train, y_train)).shuffle(10000).batch(32)
3
test_ds = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(32)

通过使用Keras模型子类 API构建tf.keras模型：

xxxxxxxxxx
15
1
class MyModel(Model):
2
  def __init__(self):
3
    super(MyModel, self).__init__()
4
    self.conv1 = Conv2D(32, 3, activation='relu')
5
    self.flatten = Flatten()
6
    self.d1 = Dense(128, activation='relu')
7
    self.d2 = Dense(10, activation='softmax')
8
​
9
  def call(self, x):
10
    x = self.conv1(x)
11
    x = self.flatten(x)
12
    x = self.d1(x)
13
    return self.d2(x)
14
​
15
model = MyModel()

选择优化器和损失函数进行训练：

xxxxxxxxxx
3
1
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
2
​
3
optimizer = tf.keras.optimizers.Adam()

选择指标（metrics）以衡量模型的损失和准确性。这些指标累积超过周期的值，然后打印整体结果。

xxxxxxxxxx
5
1
train_loss = tf.keras.metrics.Mean(name='train_loss')
2
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
3
​
4
test_loss = tf.keras.metrics.Mean(name='test_loss')
5
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')

使用tf.GradientTape训练模型：

xxxxxxxxxx
10
1
@tf.function
2
def train_step(images, labels):
3
  with tf.GradientTape() as tape:
4
    predictions = model(images)
5
    loss = loss_object(labels, predictions)
6
  gradients = tape.gradient(loss, model.trainable_variables)
7
  optimizer.apply_gradients(zip(gradients, model.trainable_variables))
8
​
9
  train_loss(loss)
10
  train_accuracy(labels, predictions)

现在测试模型：

xxxxxxxxxx
7
1
@tf.function
2
def test_step(images, labels):
3
  predictions = model(images)
4
  t_loss = loss_object(labels, predictions)
5
​
6
  test_loss(t_loss)
7
  test_accuracy(labels, predictions)

xxxxxxxxxx
15
1
EPOCHS = 5
2
​
3
for epoch in range(EPOCHS):
4
  for images, labels in train_ds:
5
    train_step(images, labels)
6
​
7
  for test_images, test_labels in test_ds:
8
    test_step(test_images, test_labels)
9
​
10
  template = 'Epoch {}, Loss: {}, Accuracy: {}, Test Loss: {}, Test Accuracy: {}'
11
  print (template.format(epoch+1,
12
                         train_loss.result(),
13
                         train_accuracy.result()*100,
14
                         test_loss.result(),
15
                         test_accuracy.result()*100))

xxxxxxxxxx
3
1
      Epoch 1, Loss: 0.13177014887332916, Accuracy: 96.06000518798828, Test Loss: 0.05814294517040253, Test Accuracy: 98.04999542236328 
2
      ...
3
      Epoch 5, Loss: 0.042211469262838364, Accuracy: 98.72000122070312, Test Loss: 0.05708516761660576, Test Accuracy: 98.3239974975586

现在，图像分类器在该数据集上的准确度达到约98％。要了解更多信息，请阅读 TensorFlow教程.。

最新版本：https://www.mashangxue123.com/tensorflow/tf2-tutorials-quickstart-advanced.html 英文版本：https://tensorflow.google.cn/beta/tutorials/quickstart/advanced 翻译建议PR：https://github.com/mashangxue/tensorflow2-zh/edit/master/r2/tutorials/quickstart/beginner.md