fix dtype not matching bug in log_prob and probs method of Distribution class

[pangyoki]

PR types

Bug fixes

PR changes

APIs

Describe

In _to_tensor method of Distribution class (refer to PR #26355 and PR #26535). Even if we want to support both float32 and float64 dtype in Distribution classes, when parameters (low and high in Uniform, loc and scale in Normal) are numpy.ndarray and dtypes are float64, we can only set dtype to be float32 using assign op to get the correspoding variable. Becase assign op doesn't support float64 when input is numpy.ndarray.

Thus, when parameters are `numpy.ndarray` with `float64` dtype, 
we will transfer them into `VarType.FP32` variable.

In log_prob and probs methods in Distribution class, the input value of these methods is a tensor. In users' view, it's reasonable that the dtype of value and parameters are same.

When dtype of parameters are `float64` in `numpy.ndarray` and dtype of `value` is `VarType.FP64`, 
it will cause error because the dtype of parameters change to `VarType.FP32`. 

The following is an example code:

import numpy as np
import paddle
from paddle.distribution import Normal

paddle.disable_static()

value_np = np.array([0.8, 0.3], dtype='float64')
value_tensor = paddle.to_tensor(value_np)  # 'float64' Tensor

loc_np = np.array([1, 2]).astype('float64')  # will be converted to 'float32' Tensor automatically
scale_np = np.array([11, 22]).astype('float64')    # will be converted to 'float32' Tensor automatically
normal = Normal(loc_np, scale_np)

lp = normal.log_prob(value_tensor)  # error !!!

We are going to let assign op support float64, but it will lose precision because Attr don't support float64 in framework.proto (refer to #26797). That is, assign op can only support float32.

• Thus, in this PR, we use cast operation to convert dtype after assign op if dtype is float64.
  If users define a Uniform distribution whose low and high are float64 numpy.ndarray, we firstly use assign op to get float32 variable. Then use cast to get float64 variable.

• What's more, probs and log_prob methods have a variable input named values. If dtype of values is different with low in Uniform or loc in Normal, it will cause error.
  To solve this dtype conflict, we cast dtype of values to be the same as that of low or loc. (in _check_values_dtype_in_probs function)

• In Doc discribtion, we add formula for entropy and kl-divergence methods. Formula for log_prob and probs have been given in doc of class, that is, the pdf (probability density function) of the distribution.

• By the way, we rewrite unittest to make it more readable.

[paddle-bot-old]

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

[zhiqiu]

Suggested change

	if str(arg_dtype) not in ['float32']:
	if str(arg_dtype) != 'float32':

[zhiqiu]

And, the positive conditions is better than the negative conditions.

[pangyoki]

Done

[zhiqiu]

Suggested change

	if str(arg_dtype) not in ['float64']:
	if str(arg_dtype) != 'float64':

[pangyoki]

Done

[zhiqiu]

Suggested change

	elif not isinstance(arg, list) and not isinstance(arg, np.ndarray):
	elif not isinstance(arg, (list, np.ndarray)):

[pangyoki]

Done

[zhiqiu]

LGTM

[jzhang533]

lgtm