fix the precision problem of test_distribution

[pangyoki]

PR types

Others

PR changes

APIs

Describe

reason for test_distribution failure

• Because assign op does not support the input of numpy.ndarray whose dtype is FP64.
  When users set FP64 numpy.ndarray as the parameters of Uniform and Normal classes. We need to use assign op to convert it to FP32 Tensor. And then use cast op to convert it to a FP64 Tensor.
  There is a loss of accuracy in this conversion.
  Refer to PR fix dtype not matching bug in log_prob and probs method of Distribution class #26767 .

• In test_distribution, compare the output of paddle and output of numpy to verify the correction.
  In Uniform(low, high), the formula to calculate the entropy is entropy(low, high) = log (high - low).
  if low and high are very close, high - low will be close to 0, and small precision loss will become large error because of using log.

solution

In the realization of the original Uniform unittest, the range of low is [-1, 1), the range of high is [-5, 5).
To avoid low and high being too close, set low in the range of [-1, 1), and set high in range of [5, 15).
What's more, add a unittest to discuss the situation that high < low.

log_prob unittest of Normal class also fails, change the tolerance from 1e-6 to 1e-4.
tolerance: 1e-6 -> 1e-4

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.