【Paddle Hackathon No.11】 #45595
【Paddle Hackathon No.11】 #45595
Conversation
| prog = paddle.static.Program() | ||
| startup_prog = paddle.static.Program() | ||
| with paddle.static.program_guard(prog, startup_prog): | ||
| input = paddle.static.data(name='input', |
There was a problem hiding this comment.
For test cases in static mode, you'd better add some cases that paddle.static.data creates data layer that has dynamic shape to ensure that it works as expected.
There was a problem hiding this comment.
I add the new test cases of checking data shape in static mode. but I'm not quite sure about it.
There was a problem hiding this comment.
I mean, you can create an data layer with shape [-1, -1] and feed it with an array with shape [3, 4] or [5, 6].
Or create a data layer with shape [-1, 3,4] and feed it with an array with shape [7, 3, 4]
I the rank and all the static size matches, it is compatible.
There was a problem hiding this comment.
- Weight should be for each class, instead of each example. The main logic is different than the design in RFC;
- use paddle.shape(tensor) to get actual shape so it can work with dynamic shape in static mode;
- add test cases for dynamic shape in static mode;
- The documentation should be modified.
There was a problem hiding this comment.
The CI system says
2022-09-26 16:20:48 0. It is recommended to use 'np.testing.assert_allclose' and 'np.testing.array_equal' instead of 'self.assertTrue(np.allclose(...))' and 'self.assertTrue(np.array_equal(...))'.
Also, I suggest a simpler implementation to pick N values from weight of shape (C,). See the comment above.
Done, change 'self.asserrTrue' to 'np.testing.assert_allclose' |
python/paddle/nn/functional/loss.py
Outdated
| reduction='mean', | ||
| name=None): | ||
| r""" | ||
| Measures a multi-class classification hinge loss (margin-based loss) between input :math:`input` (a 2D mini-batch `Tensor`, in shape (N, C), |
There was a problem hiding this comment.
参考paddle.nn.MultiMarginLoss做修改吧
python/paddle/nn/layer/loss.py
Outdated
|
|
||
| label: 1-D Tensor, the shape is [N,]. | ||
|
|
||
| output: scalar. If :attr:`reduction` is ``'none'``, then same shape as the input. |
There was a problem hiding this comment.
好像不是一个意思？input不是2-D tensor吗？中文文档中的返回是一个1-D tensor？
There was a problem hiding this comment.
应该是和 label 同一维度，能否合入后，我再发起 pr 修改
There was a problem hiding this comment.
文档修改可以在commit后加上 ; test=document_fix 跳过代码检查的CI
| margin: float = 1.0, | ||
| weight=None, | ||
| reduction="mean", | ||
| name=None): |
There was a problem hiding this comment.
in order to have rfc same with code, need to modify paddle.nn.MultiMarginLoss( p:int = 1, margin:float=1.0, weight (Tensor,可选), reduction(str,可选), name=None) -> None: in rfc 命名与参数设计
|
@yangguohao 啊哦，示例代码挂了，重新修改一下吧，可以采用 |
该问题应由 #46712 导致，最近的只要修改了文档的 PR 均存在此问题，#46944 merge 后应可解决本问题 @yangguohao 需要 #46944 merge 后，重新 merge 下 develop |
PR types
New features
PR changes
APIs
Describe
解决 issues #44073 (comment)
添加了paddle.nn.MultiMarginLoss以及paddle.nn.functional.multi_margin_loss的函数
设计文档 PaddlePaddle/community#171
中文文档 PaddlePaddle/docs#5262