[Dy2stat]support Python3 type annotation

[0x45f]

PR types

Others

PR changes

Others

Describe

Dy2stat type annotation

背景
python3支持type hint语法，动转静模块中可以借助type hint语法进行变量类型分析。

功能描述
python3中的type hint语法，比如赋值语句a: float = 1.0和函数参数func(a: int, b: float = 1)。对于下面的代码forward函数虽然将texts标示为paddle.Tensor类型，但是由于动转静模块并未支持python3的type hint语法，无法将texts识别为一个tensor，for语句动转静处理list时会出错。在此PR之后可以将函数参数texts识别为一个tensor。

import paddle 
from paddle.jit import to_static
from paddle import nn
import paddle.nn.functional as F

paddle.set_device('cpu')

class SDMGRHead(nn.Layer):
    def __init__(self,):
        super(SDMGRHead, self).__init__()
        self.name = "name"

        #self.edge_ = nn.Linear(5, 256)
        num_chars = 92
        node_input = 32
        self.edge_embed = nn.Embedding(num_chars, node_input, 0)
    
    def forward(self, texts: paddle.Tensor):
        node_nums, char_nums = [], []
        # import pdb; pdb.set_trace()
        for text in texts:
            # node_nums.append(text.shape[0])
            char_nums.append(paddle.sum((text > -1).astype('int32'), axis=-1))

        max_num = max([char_num.max() for char_num in char_nums])
        #max_num = paddle.max(paddle.concat(char_nums)).numpy()[0]
        all_nodes = paddle.concat([
            paddle.concat(
                [text, paddle.zeros(
                    (text.shape[0], max_num - text.shape[1]))], -1)
            for text in texts
        ])
        temp = paddle.clip(all_nodes, min=0).astype(int)
        embed_nodes = self.edge_embed(temp)
        return embed_nodes

texts = paddle.randn((1, 300, 300)).astype('float32') + 3
model = SDMGRHead()
res = model(texts)
print(res.shape)
infer_shape = [paddle.static.InputSpec(shape=[1, 300, 300], dtype="float32")]
model = to_static(model, input_spec=infer_shape)
paddle.jit.save(model, "./debug/inference")
print("debug: save_inference success!")

测试
下面是两个测试样例，以及StaticAnalysisVisitor模块所识别出的变量类型。

def func_to_test3():
    a = 1
    b = 3.0
    c = a * b
    d = True + c
    e = a < b
    f = 9 * (a * 4)
    g = "dddy"
    h = None
    i = False
    j = None + 1
    k: float = 1.0
    l: paddle.Tensor = paddle.to_tensor([1, 2])
result_var_type3 = {
    'a': {NodeVarType.INT},
    'b': {NodeVarType.FLOAT},
    'c': {NodeVarType.FLOAT},
    'd': {NodeVarType.FLOAT},
    'e': {NodeVarType.BOOLEAN},
    'f': {NodeVarType.INT},
    'g': {NodeVarType.STRING},
    'h': {NodeVarType.NONE},
    'i': {NodeVarType.BOOLEAN},
    'j': {NodeVarType.UNKNOWN},
    'k': {NodeVarType.FLOAT},
    'l': {NodeVarType.PADDLE_RETURN_TYPES}
}

def func_to_test7(a: int, b: float, c: paddle.Tensor, d: float='diff'):
    a = True
result_var_type7 = {
    'a': {NodeVarType.BOOLEAN},
    'b': {NodeVarType.FLOAT},
    'c': {NodeVarType.TENSOR},
    'd': {NodeVarType.STRING}
}

[paddle-bot-old]

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

[Aurelius84]

Great work~

[Aurelius84]

这里只会对类似 x: float = 2.1 作处理，对于类似self.x: float = 2.1 是不会处理？ 因为self.x是一个gast.Attribute，而不是一个gast.Name.

此PR 若是不支持的，这里记一个TODO，需要确认下是否有类似非gats.Name的场景和需求

[0x45f]

在之前代码中的Assign node也没有支持类似self.x= 2.1赋值语句，所以在AnnAssign node也暂时不支持self.x: float = 2.1类似语句。后续可以确认是否需要支持此类赋值语句。

[Aurelius84]

to_variable 接口已经被弃用了，应该避免使用旧的接口

[0x45f]

已修改使用to_tensor接口