【Hackathon 5th No.33】为 Paddle 新增 atleast_1d / atleast_2d / atleast_3d API -part #58323
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
Sorry to inform you that 2395977's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
python/paddle/tensor/manipulation.py
Outdated
| name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. | ||
|
|
||
| Note: | ||
| ``int8``, ``uint8``, ``complex64``, ``complex128`` are not supported in static graph mode. |
There was a problem hiding this comment.
这几个方法里面用到了 reshape 将 0-dim 转为 n-dim，当时 rfc 的时候没想到 reshape 的动/静态图支持的不一样，静态图不支持这几个类型：
def reshape(x, shape, name=None):
if in_dynamic_mode():
...
else:
check_variable_and_dtype(
x,
'x',
[
'float16',
'float32',
'float64',
'int16',
'int32',
'int64',
'bool',
'uint16',
],
'reshape',
)所以这里单独将这几个类型单拎出来了 ～
另外，bool 应该也支持，当时 rfc 的时候没有提到 bool 类型，是否也加进去？
There was a problem hiding this comment.
reshape 的 check_variable_and_dtype 是否可以加上这几个类型呢？如果可以的话，需要单独提一个PR支持静态图下reshape的 int8, uint8, complex64, complex128类型。
There was a problem hiding this comment.
看了一下底层 c++ 算子的实现，没找到 reshape 对于这几种数据类型有单独处理。
尝试修改 paddle/tensor/manipulation.py :
def reshape(x, shape, name=None):
if in_dynamic_mode():
...
elif in_pir_mode():
...
else:
check_variable_and_dtype(
x,
'x',
[
'float16',
'float32',
'float64',
'int16',
'int32',
'int64',
'bool',
'uint16',
'int8',
'uint8',
'complex64',
'complex128',
],
'reshape',
)简单验证：
In [1]: import numpy as np
...: import paddle
...:
...: paddle.enable_static()
...: with paddle.static.program_guard(paddle.static.Program()):
...: feed = {}
...: x = paddle.static.data('x', [12], 'int8')
...: feed['x'] = np.arange(12).astype('int8')
...:
...: out = paddle.reshape(x, [2, 6])
...: exe = paddle.static.Executor(paddle.CUDAPlace(0))
...: res = exe.run(feed=feed, fetch_list=[out])
...:
...: print(res)
...:
[array([[ 0, 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10, 11]], dtype=int8)]- 未修改之前，会报错 TypeError
- 修改之后，能够正常运行
另外，in_pir_mode 这里先不修改吧 ～ 我试了一下 os.environ['FLAGS_enable_new_ir_in_executor'] = 'True'，接口会挂掉 ... ...
那我这边单独提 PR 再逐个类型验证一下吧 ～
Update 20231104
|
|
@megemini 可以merge develop重跑CI了 |
| from .manipulation import atleast_1d # noqa: F401 | ||
| from .manipulation import atleast_2d # noqa: F401 | ||
| from .manipulation import atleast_3d # noqa: F401 |
There was a problem hiding this comment.
Shall we need to support the usage of x.atleast_1d(), x.atleast_2d(), x.atleast_3d() when x is a tensor ? If so, add them to the list of tensor_method_func below
test/legacy_test/test_atleast_nd.py
Outdated
| ), | ||
| ) | ||
| class TestAtleastErrorCombineInputs(BaseTest): | ||
| """test combine inputs, like: `at_leastNd((x, y))`, where paddle treats like numpy, not pytorch""" |
There was a problem hiding this comment.
please delete not pytorch
There was a problem hiding this comment.
cc @zhwesky2010 ：atleast_xd 处理语义在遇到tuple/list时和numpy对齐，和pytorch有差异，后续在代码转换时可以关注下
There was a problem hiding this comment.
cc @zhwesky2010 ：atleast_xd 处理语义在遇到tuple/list时和numpy对齐，和pytorch有差异，后续在代码转换时可以关注下
主要区别是：
- paddle/numpy -> paddle.atleast_1d((x, y)) 将 (x, y) 作为一个整体输入，并转换为 tensor （此处需要注意 broadcast）
- pytorch -> torch.atleast_1d((x, y)) 将 (x, y) 作为两个单独的输入 x 和 y
个人感觉 torch 这种处理方式是有歧义的，如 torch.atleast_1d((x, y)) 和 torch.atleast_1d(x, y) 效果是一样的：
In [4]: y = torch.tensor(0.3)
In [5]: x = torch.tensor(0.3)
In [6]: torch.atleast_1d(x, y)
Out[6]: (tensor([0.3000]), tensor([0.3000]))
In [7]: torch.atleast_1d((x, y))
Out[7]: (tensor([0.3000]), tensor([0.3000]))所以，根据本次设计接口的签名: paddle.atleast_1d(*inputs)，将 paddle.atleast_1d((x, y)) 作为一个输入，而 paddle.atleast_1d(x, y) 是两个输入。
请悉知 ～
Update 20231115
|
|
@megemini 请补充对应的中文文档 |
luotao1
changed the title
|
@megemini 顺师傅能修一下单测随机挂么？ 单测 : test_atleast_nd, PR: 57785,CI : PR-CI-Mac-Python3,url : https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/9541163 
|
… API -part (PaddlePaddle#58323) * [Init] add atleast api * [Add] add atleast test * [Fix] import atleast * [Change] test_atleast.py to test_atleast_nd.py and add bool data type test * [Update] update dtype supports and unittest * [Fix] dtype error unittest * [Change] static test with test_with_pir_api * [Add] atleast_Nd as tensor method
PR types
New features
PR changes
APIs
Description
【Hackathon 5th No.33】为 Paddle 新增 atleast_1d / atleast_2d / atleast_3d API
RFC： PaddlePaddle/community#679
涉及文件：
python/paddle/__init__.py将 API 暴露出来python/paddle/tensor/__init__.py将 API 暴露出来python/paddle/tensor/manipulation.py实现 APItest/legacy_test/test_atleast.py单元测试本地测试情况：
请评审！