[Typing][A-17] Add type annotations for conv layers #65183
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
python/paddle/nn/layer/conv.py
Outdated
| filter_elem_num = num_channels * np.prod(filter_size) | ||
| std = (2.0 / filter_elem_num) ** 0.5 | ||
| return Normal(0.0, std) | ||
|
|
||
|
|
||
| def _reverse_repeat_list(t, n): | ||
| def _reverse_repeat_list(t: Sequence[int], n: int) -> list: |
There was a problem hiding this comment.
这里的 t 应该不只限于 int ～ 可以试一下 type var ~
def _reverse_repeat_list(t: Sequence[_T], n: int) -> list[_T]:
python/paddle/nn/layer/conv.py
Outdated
| in_channels: Any, | ||
| out_channels: Any, | ||
| kernel_size: Any, | ||
| transposed: Any, | ||
| dims: Any, | ||
| stride: int | IntSequence = 1, | ||
| padding: int | IntSequence = 0, | ||
| padding_mode: str = 'zeros', | ||
| output_padding: int | IntSequence = 0, | ||
| dilation: int | IntSequence = 1, | ||
| groups: int = 1, | ||
| weight_attr: Any | None = None, | ||
| bias_attr: Any | None = None, | ||
| data_format: str = "NCHW", | ||
| ) -> None: |
There was a problem hiding this comment.
尽量不要用 Any ～
另外，data_format 在 _typing 模块里面应该有公用类型 ～
There was a problem hiding this comment.
不要直接从 paddlepaddle-stubs copy，要思考，_ConvNd 明显没有标注
python/paddle/nn/layer/conv.py
Outdated
| from ..._typing import ( | ||
| DataLayoutND, | ||
| DataLayout1D, | ||
| DataLayout2D, | ||
| DataLayout3D, | ||
| IntSequence, | ||
| ShapeLike, | ||
| ) | ||
|
|
||
| PaddingSizeStr: TypeAlias = Literal["valid", "same"] | ||
| PaddingMode: TypeAlias = Literal["zeros", "reflect", "replicate", "circular"] |
There was a problem hiding this comment.
以及下面的 from paddle import ParamAttr 放到 if TYPE_CHECKING: 里面吧 ～
2024-06-25 16:49:54 ImportError: cannot import name 'ParamAttr' from partially initialized module 'paddle' (most likely due to a circular import) (/paddle/build/python/paddle/__init__.py)这里的报错也跟这个有关 ～
另外，这里依赖的 PaddingSizeStr 可以等 #65197 合入后，从 python/paddle/nn/functional/common.py 引入 ～
TYPE_CHECKING 里面再加一个 from paddle import Tensor 吧 ～ 后面就不用 paddle.Tensor 而是用 Tensor ，大家统一一下 ～
python/paddle/nn/layer/conv.py
Outdated
| filter_elem_num = num_channels * np.prod(filter_size) | ||
| std = (2.0 / filter_elem_num) ** 0.5 | ||
| return Normal(0.0, std) | ||
|
|
||
|
|
||
| def _reverse_repeat_list(t, n): | ||
| def _reverse_repeat_list(t: Sequence[int | str], n: int) -> list: |
There was a problem hiding this comment.
可以尝试
_T = TypeVar("_T")
def _reverse_repeat_list(t: Sequence[_T], n:int) -> list[_T]:
"""Reverse the order of `t` and repeat each element for `n` times.
This can be used to translate padding arg used by Conv and Pooling modules
to the ones used by `F.pad`.
"""
return [x for x in reversed(t) for _ in range(n)]
python/paddle/nn/layer/conv.py
Outdated
| weight_attr: Any | None = None, | ||
| bias_attr: Any | None = None, |
python/paddle/nn/layer/conv.py
Outdated
| stride: int | IntSequence = 1, | ||
| padding: int | IntSequence | PaddingSizeStr = 0, | ||
| dilation: int | IntSequence = 1, |
There was a problem hiding this comment.
根据不同的 Conv ，这里的输入可能不一样 ～ 参考 #65191
后面几个也是 ～
|
这个 PR 后续我来推进 |
SigureMo
changed the title
| @@ -63,7 +63,7 @@ | |||
| ] | |||
| _DropoutMode: TypeAlias = Literal['upscale_in_train', 'downscale_in_infer'] | |||
| _PaddingTensorMode: TypeAlias = Literal[ | |||
There was a problem hiding this comment.
为啥要加 zeros? 我记得当时单独在 conv 里面加了一个
_ConvPaddingMode: TypeAlias = Literal[
"zero", "reflect", "replicate", "circular"
]_PaddingTensorMode 是 constant ... ，_ConvPaddingMode 是 zeros ... ～
PR Category
User Experience
PR Types
Improvements
Description
类型标注：
Related links
@SigureMo @megemini