Add eager string tensor

[joey12300]

PR types

New features

PR changes

Others

Describe

接口描述

新增core.eager.StringTensor接口，目前主要添加了6个初始化方法以及numpy方法，结构如下：

class StringTensor:
   def __init__(): 
       # case 1
       pass
   def __init__(dims: vector<int>, name: std::string):
       # case 2
       pass
   def __init__(value: ndarray, name: std::string): #去除place, zero_copy参数
       # case 3
       pass
   def __init__(value: ndarray):
       # case 4
       pass
   def __init__(tensor: Tensor):
       # case 5
       pass
   def __init__(tensor: Tensor, name: std::string): #去除place参数
       # case 6
       pass
   def numpy():
       # 返回Unicode类型的numpy array
       pass

core.eager.StringTensor的初始化方法接口设计参考core.eager.Tensor的接口，并根据StringTensor的需求删除了dtype，persistable, stop_gradient三个参数输入。目前StringTensor暂时只对外暴露CPU上的功能，所以在接口中也删除place参数。至于zero_copy参数，由于StringTensor的内存分布与numpy的string array内存分布不一致，无法直接使用numpy的string array初始化StringTensor，所以这里删除了zero_copy。core.eager.StringTensor不支持使用FrameworkTensor初始化。

使用示例

import numpy as np
import paddle
from paddle.fluid import core

A_arr= np.array([
  ["15.4寸笔记本的键盘确实爽，基本跟台式机差不多了，蛮喜欢数字小键盘"],  # From ChnSentiCorp
  ["One of the very best Three Stooges shorts ever."]])  # From IMDB

ST1 = core.eager.StringTensor()  # case 1
ST2 = core.eager.StringTensor([2, 3], "ST2")  # case 2
ST3 = core.eager.StringTensor(A_arr, "ST3")  # case 3
ST4 = core.eager.StringTensor(A_arr)  # case 4
ST5 = core.eager.StringTensor(ST4)  # case 5

string_tensors = [ ST1, ST2, ST3, ST4, ST5]
for i, tensor in enumerate(string_tensors):
    lower_tensor = core.eager.ops.final_state_strings_lower(tensor, True)
    upper_tensor = core.eager.ops.final_state_strings_upper(tensor, True)
    print("case {}: Tensor={}, name={}, shape={}, dtype={}, place={}, numpy={}".format(
      i, tensor, tensor.name, tensor.shape, tensor.dtype, tensor.place, tensor.numpy()))
    print("case {}: LowerTensor={}, name={}, shape={}, dtype={}, place={}, numpy={}".format(
      i, lower_tensor, lower_tensor.name, lower_tensor.shape, lower_tensor.dtype, lower_tensor.place, lower_tensor.numpy()))
    print("case {}: UpperTensor={}, name={}, shape={}, dtype={}, place={}, numpy={}".format(
      i, upper_tensor, upper_tensor.name, upper_tensor.shape, upper_tensor.dtype, upper_tensor.place, upper_tensor.numpy()))

[paddle-bot-old]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[JiabinYang]

LGTM