Created by: pangyoki

PR types

Others

PR changes

APIs

Describe

In Paddle 2.0-beta, PR #26006 add a new default_dtype API. This PR is used to support global default dtype for distribution related APIs.

Distribution related APIs include:

- op：
    - paddle.tensor.random.bernoulli
    - paddle.tensor.random.gaussian_random
    - paddle.tensor.random.standard_normal
    - paddle.tensor.random.normal
    - paddle.tensor.random.uniform
    - paddle.tensor.random.randn
    - paddle.tensor.random.rand
    - paddle.tensor.random.randint
    - paddle.tensor.random.randperm

- sample method of distribution related classes
    - paddle.distribution.Normal
    - paddle.distribution.Uniform

They are divided into the following categories:

• have dtype parameter and the default value is float32, including:

- paddle.tensor.random.gaussian_random
- paddle.tensor.random.standard_normal
- paddle.tensor.random.uniform
- paddle.tensor.random.randn
- paddle.tensor.random.rand

method: Set None to dtype parameter. if dtype is None, get the value of dtype from global dtype. Because global dtype support [float16, float32, float64], but these distribution related APIs only support [float32, float64]. So if global dtype is float16, remind users that the dtype has error.

• have dtype parameter and the default value is int, including:

- paddle.tensor.random.randint
- paddle.tensor.random.randperm

Can't use global dtype for global dtype not supporting int.

• don't have dtype parameter and have input tensor, including:

- paddle.tensor.random.bernoulli

Be consistent with Pytorch. get the output value of dtype from dtype of input tensor. Don't use global dtype.

• Call other distritbuion APIs, including:

- paddle.tensor.random.normal
- sample method in paddle.distribution.Normal
- sample method in paddle.distribution.Uniform

Let the dtype of uniform be consistent with the dtype of low parameter and high parameter. Let the dtype of normal be consistent with the dtype of mean parameter and std parameter. Don't use default dtype.