在 TensorFlow 中禁用急切执行时如何计算均方误差？

Question

When calculating MSE using tensorflow, I get the error AttributeError: 'Tensor' object has no attribute 'numpy' The reason is that I need to disable eager execution ( tf.disable_eager_execution() ). Question: How to calculate mean square error when eager execution is disabled in TensorFlow? The code looks something like this (I'm using the latest version of tenorflow):

tf.disable_eager_execution()

mse = tf.keras.losses.MeanSquaredError()
MSE = mse(y_true, y_prediction).numpy()

Answer 1

You can't and shouldn't. When eager execution is disabled, the calculations and objects are leaving Python. The goal of this is to train a model with an optimized backend rather than "slow" Python. As a side effect, the objects and values aren't accessible to Python. This is fine when you train a model, which you would be able to run in Python. However, this doesn't work if you want to monitor every operation.

If you want to access tensor values, just enable eager execution.

Answer 2

You need to create a session. This should work (tested on TF 2.5):

tf.compat.v1.disable_eager_execution()

mse = tf.keras.losses.MeanSquaredError()
MSE = mse(y_true, y_prediction)

with tf.compat.v1.Session().as_default():
    MSE.eval()

Though I concur with Nicolas that debugging is better (and more easily) done in eager mode.