名称作用域

当处理更复杂的模型（如神经网络）时，该图可以很容易地与数千个节点混淆。 为了避免这种情况，您可以创建名称作用域来对相关节点进行分组。 例如，我们修改以前的代码来定义名为loss的名称作用域内的错误和mse操作：

  1. with tf.name_scope("loss") as scope:
  2.     error = y_pred - y
  3.     mse = tf.reduce_mean(tf.square(error), name="mse")

在作用域内定义的每个op的名称现在以loss/为前缀：

  1. >>> print(error.op.name)
  2. loss/sub
  3. >>> print(mse.op.name)
  4. loss/mse

在 TensorBoard 中，mse和error节点现在出现在loss命名空间中，默认情况下会出现崩溃（图 9-5）。

名称作用域 - 图1

完整代码

  1. import numpy as np
  2. from sklearn.datasets import fetch_california_housing
  3. import tensorflow as tf
  4. from sklearn.preprocessing import StandardScaler
  6. housing = fetch_california_housing()
  7. m, n = housing.data.shape
  8. print("数据集:{}行,{}列".format(m,n))
  9. housing_data_plus_bias = np.c_[np.ones((m, 1)), housing.data]
  10. scaler = StandardScaler()
  11. scaled_housing_data = scaler.fit_transform(housing.data)
  12. scaled_housing_data_plus_bias = np.c_[np.ones((m, 1)), scaled_housing_data]
  14. from datetime import datetime
  16. now = datetime.utcnow().strftime("%Y%m%d%H%M%S")
  17. root_logdir = r"D://tf_logs"
  18. logdir = "{}/run-{}/".format(root_logdir, now)
  20. n_epochs = 1000
  21. learning_rate = 0.01
  23. X = tf.placeholder(tf.float32, shape=(None, n + 1), name="X")
  24. y = tf.placeholder(tf.float32, shape=(None, 1), name="y")
  25. theta = tf.Variable(tf.random_uniform([n + 1, 1], -1.0, 1.0, seed=42), name="theta")
  26. y_pred = tf.matmul(X, theta, name="predictions")
  28. def fetch_batch(epoch, batch_index, batch_size):
  29.     np.random.seed(epoch * n_batches + batch_index)  #  not shown in the book
  30.     indices = np.random.randint(m, size=batch_size)  #  not shown
  31.     X_batch = scaled_housing_data_plus_bias[indices] #  not shown
  32.     y_batch = housing.target.reshape(-1, 1)[indices] #  not shown
  33.     return X_batch, y_batch
  35. with tf.name_scope("loss") as scope:
  36.     error = y_pred - y
  37.     mse = tf.reduce_mean(tf.square(error), name="mse")
  39. optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
  40. training_op = optimizer.minimize(mse)
  42. init = tf.global_variables_initializer()
  44. mse_summary = tf.summary.scalar('MSE', mse)
  45. file_writer = tf.summary.FileWriter(logdir, tf.get_default_graph())
  47. n_epochs = 10
  48. batch_size = 100
  49. n_batches = int(np.ceil(m / batch_size))
  51. with tf.Session() as sess:
  52.     sess.run(init)
  54.     for epoch in range(n_epochs):
  55.         for batch_index in range(n_batches):
  56.             X_batch, y_batch = fetch_batch(epoch, batch_index, batch_size)
  57.             if batch_index % 10 == 0:
  58.                 summary_str = mse_summary.eval(feed_dict={X: X_batch, y: y_batch})
  59.                 step = epoch * n_batches + batch_index
  60.                 file_writer.add_summary(summary_str, step)
  61.             sess.run(training_op, feed_dict={X: X_batch, y: y_batch})
  63.     best_theta = theta.eval()
  65. file_writer.flush()
  66. file_writer.close()
  67. print("Best theta:")
  68. print(best_theta)

名称作用域 - 图2