[0-size Tensor Job2 No.51] Add 0-size Tensor support for paddle.multiplex [fluid_ops] #74212
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
/re-run all-failed |
Codecov Report❌ Patch coverage is
❌ Your patch status has failed because the patch coverage (0.00%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## develop #74212 +/- ##
==========================================
Coverage ? 0.00%
==========================================
Files ? 1
Lines ? 1
Branches ? 0
==========================================
Hits ? 0
Misses ? 1
Partials ? 0 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
|
/re-run all-failed |
|
@DanielSun11 CI已完成需要review |
| ins.size(), | ||
| errors::PreconditionNotMet( | ||
| "index exceeds the number of candidate tensors.")); | ||
| if (ins[k]->numel() == 0) continue; |
There was a problem hiding this comment.
所有的tensor都不为0-size或者都为0-size，这里处理的是所有的tensor numel都为0，都会跳过
There was a problem hiding this comment.
也就是说修改后的paddle.multiplex只能支持输入全部为0 size或者输入全部为非0 size的情况吗？如果是这样的，infermeta和符号推导就没有必要改了。https://www.paddlepaddle.org.cn/documentation/docs/zh/3.0-beta/api/paddle/multiplex_cn.html#multiplex 规定了所有输入的shape必须相同，如果输入为0 size则其所有输入的shape也要求相同。只需要在kernel中判断如果out的numel为0，就直接return，就可以了吧？shape不同的为非法case
There was a problem hiding this comment.
也就是说修改后的paddle.multiplex只能支持输入全部为0 size或者输入全部为非0 size的情况吗？如果是这样的，infermeta和符号推导就没有必要改了。https://www.paddlepaddle.org.cn/documentation/docs/zh/3.0-beta/api/paddle/multiplex_cn.html#multiplex 规定了所有输入的shape必须相同，如果输入为0 size则其所有输入的shape也要求相同。只需要在kernel中判断如果out的numel为0，就直接return，就可以了吧？shape不同的为非法case
输入可以包含0-size和非0-size，输出取非0-size的shape，但是index选择的时候，选择的是非0-size项，如果都是选择0-size项，kernel结尾要Resize，这里看下torch的规则大概可以明白，构造temp时都是非0-size或者都是0-size，但是输入可以是不同的
文档描述shape是一致，但是0-size特例也可能出现，所以这里加了处理
There was a problem hiding this comment.
import paddle
x1 = paddle.randn([3, 3],dtype=paddle.float32)
x2 = paddle.randn([3, 0],dtype=paddle.float32)
print(x1)
print(x2)
index = paddle.to_tensor([[0],[1]],dtype=paddle.int32)
z = paddle.multiplex([x1, x2], index)
print(z)
请验证下这个代码的输出
There was a problem hiding this comment.
那还是按简单的处理，规则复杂也不容易维护，输入shape都一致 测试用例修改PR PFCCLab/PaddleAPITest#444
|
/re-run all-failed |
| ins.size(), | ||
| errors::PreconditionNotMet( | ||
| "index exceeds the number of candidate tensors.")); | ||
| if (ins[k]->numel() == 0) continue; |
There was a problem hiding this comment.
import paddle
x1 = paddle.randn([3, 3],dtype=paddle.float32)
x2 = paddle.randn([3, 0],dtype=paddle.float32)
print(x1)
print(x2)
index = paddle.to_tensor([[0],[1]],dtype=paddle.int32)
z = paddle.multiplex([x1, x2], index)
print(z)
请验证下这个代码的输出
4 participants
PR Category
Operator Mechanism
PR Types
Bug fixes
Description
输入shape一致都为0-size，判断out numel返回
PaddleAPITest中为自定义规则torch.stack
https://github.com/PFCCLab/PaddleAPITest/blob/81e9f10d78ec53b8a56684b75bcdcf108f38a43f/tester/paddle_to_torch/rules.py#L3795
增加单测
PaddleAPITest测试通过，错误为 torch error