tf.metrics.mean返回值含义解析：为何两次求值结果不同？

Let's break down exactly what's happening with your code and why you're seeing those seemingly inconsistent results.

Core Background: What tf.metrics.mean Returns #

When you call tf.metrics.mean(test_tensor), it returns a tuple of two distinct TensorFlow objects:

1. mean (first element): A tensor representing the cumulative average of all values processed so far. This value relies on two internal local variables: a running total of all values, and a count of how many values have been added. Every time you run the update_op, these variables get updated.
2. update_op (second element): An operation that does two key things:
   • Calculates the average of the current input tensor (in your case, this is always 3.5 since your input is a constant tensor).
   • Adds the sum of the input tensor to the running total, and increments the count by the number of elements in the input.
   • Finally, it returns the average of the current input (not the cumulative average).

Why Your Output Looks the Way It Does #

Let's walk through each step of your execution to clarify:

1. First sess.run(test_mean):

   • TensorFlow reads the mean tensor first (before update_op runs). Initially, the running total and count are both 0, so mean returns 0.0.
   • Then update_op executes: it calculates the input's average (3.5), updates the running total to 21 (sum of your 6 elements) and count to 6. It returns the input's average (3.5).
   • Result: (0.0, 3.5)

2. Second to fourth sess.run(test_mean):

   • Each time you run the tuple, TensorFlow reads the current cumulative mean before executing update_op. The mean value you see is the cumulative average from all previous updates, not including the current one. This is why it gradually approaches 3.5: each update_op adds another full batch of values to the running state, and you're seeing the state before that latest addition.
   • Results: (1.75, 3.5), (2.3333333, 3.5), (2.625, 3.5)

3. sess.run(test_mean[0]):

   • By this point, you've run update_op four times. The cumulative total and count have been fully updated to reflect all four batches, so the cumulative average is exactly 3.5. When you read mean directly, you're getting the latest, fully updated cumulative value.
   • Result: 3.5

4. sess.run(test_mean[1]):

   • The update_op always recalculates the average of your fixed input tensor, so it will always return 3.5, no matter how many times you run it.
   • Result: 3.5

Key Takeaway #

• Running the full tuple test_mean gives you the old cumulative state before updating it, plus the current input's average.
• Running test_mean[0] alone reads the latest, fully updated cumulative average.
• test_mean[1] always returns the average of the current input, independent of the cumulative state.

内容的提问来源于stack exchange，提问作者quant