[0-size Tensor No.353] Add 0-size Tensor support for unflatten API. by luyl975 · Pull Request #73986 · PaddlePaddle/Paddle

luyl975 · 2025-07-11T02:44:49Z

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

a.错误分析
在PaddleAPITest report/0size_tensor中检索paddle.unflatten的错误日志。

[paddle error] paddle.unflatten(x=Tensor([4, 0, 16],"float32"), axis=0, shape=tuple(-1,), ) 
 (InvalidArgument) can not reshape 4, 0, 16 to -1, 0, 16, because the unspecified dimension 0 can be any number and is ambiguous
  [Hint: Expected unk_dim_idx == -1, but received unk_dim_idx:0 != -1:-1.] (at /paddle/paddle/phi/infermeta/unary.cc:2209)

定位至源代码（由于github其它分支的合并，具体行数不为2209行），发现报错函数为/paddle/paddle/phi/infermeta/unary.cc中的ValidateShape，这个函数的作用是计算结果张量的形状。

分析该函数逻辑，发现该函数传入的内容包括2项，即shape与in_dims。其中，in_dims是传入原张量的形状，shape（后记为vshape与paddle.unflatten中的shape参数区分）是将希望展开的形状嵌入原张量的形状。用以下代码的执行为例：

paddle.unflatten(paddle.randn([4, 6, 2]), axis=2, shape=(-1,1))

此时传入ValidateShape的in_dims为[4, 6, 2]，vshape为[4, 6, -1, 1]，其中shape中的值不会被做任何替换。
在执行至错误代码前，会构造一个“std::vector<int64_t> output_shape(shape.size(), 0)”，表示结果张量的形状，并根据一定规则填充vshape中的值（如发现-1个数大于1时报错，因为不定的位置只能有1个）。
错误代码报错的原因在于：in_dims中含0且执行至此处时，变量unk_dim_idx的值不为-1。
其中，unk_dim_idx记录的是shape中-1的位置。会对于这一情况报错的原因在于，后续代码中需要确定-1这一不定的值，确定的方法是用“in_dims中各个元素的乘积”除以“shape中-1外各个元素的乘积”，即源代码中的“output_shape[unk_dim_idx] = in_size / capacity“。
当输出0-size张量时，这一除法显然无法成立，因此选择在vshape中同时含有-1和0时报错。

b.错误解决
对于vshape含有-1且in_dims含有0运行至报错位置时的情况可以分三种情况讨论：
1、 vshape中与in_dims中0的个数相同
出现这种情况说明vshape是由in_dims中一个非0数被替换为含-1的shape构成，将in_dims中各个非0数之积除以vshape中各个非-1与0的乘积即可。
2、 vshape中0的个数少于in_dims中0的个数
出现这种情况说明vshape是由in_dims中的0替换为含-1的shape构成，则此时vshape中-1无法确定具体的值，应当报错，可使用源代码相关步骤，不必进行额外修改。
3、 vshape中0的个数多于in_dims中0的个数
出现这种情况可以分两种情况讨论：
情况一：in_dims中非0数被shape替换，且shape中同时含有-1和0，此时参考torch，应当报错：

情况二：in_dims中的0被shape替换，且shape中同时含有-1和至少2个0，此时-1无法确定具体的值，应当报错，可使用源代码相关步骤，不必进行额外修改。

Codecov Report

❌ Patch coverage is 95.65217% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@bfa8da2). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddle/phi/infermeta/unary.cc	95.65%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #73986   +/-   ##
==========================================
  Coverage           ?   95.65%           
==========================================
  Files              ?        1           
  Lines              ?       23           
  Branches           ?        0           
==========================================
  Hits               ?       22           
  Misses             ?        1           
  Partials           ?        0

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

luyl975 · 2025-07-18T16:00:36Z

/re-run all-failed

paddle/phi/infermeta/unary.cc

luyl975 · 2025-07-21T05:53:46Z

删掉这些无关的文件

已删除

luyl975 · 2025-07-21T17:23:59Z

/re-run all-failed

luyl975 · 2025-07-22T02:19:04Z

/re-run all-failed

DanielSun11 · 2025-07-22T06:59:04Z

paddle/phi/infermeta/unary.cc

+    if (unk_dim_idx != -1) {
+      size_t in_dims_zero_cnt = 0;
+      for (size_t i = 0; i < in_dims_vec.size(); ++i)
+        if (in_dims_vec[i] == 0) in_dims_zero_cnt++;
+      if (shape_zero_cnt == in_dims_zero_cnt) {
+        int64_t in_dims_pdt = 1;
+        int64_t shape_pdt = 1;
+        for (size_t i = 0; i < shape.size(); ++i)
+          if (shape[i] != 0 && shape[i] != -1) shape_pdt *= shape[i];
+        for (size_t i = 0; i < in_dims_vec.size(); ++i)
+          if (in_dims_vec[i] != 0 && in_dims_vec[i] != -1)
+            in_dims_pdt *= in_dims_vec[i];
+        output_shape[unk_dim_idx] = in_dims_pdt / shape_pdt;
+        PADDLE_ENFORCE_EQ(
+            output_shape[unk_dim_idx] * shape_pdt,
+            in_dims_pdt,
+            common::errors::InvalidArgument(
+                "The 'shape' attribute in ReshapeOp is invalid. "
+                "The input tensor X'size must be divisible by known "
+                "capacity of 'shape'. "
+                "But received X's shape = [%s], "
+                "'shape' is [%s].",
+                in_dims,
+                common::make_ddim(shape)));
+        return common::make_ddim(output_shape);
+      } else if (shape_zero_cnt > in_dims_zero_cnt) {
+        int64_t in_dims_pdt = 1;
+        int64_t shape_pdt = 1;
+        for (size_t i = 0; i < shape.size(); ++i)
+          if (shape[i] != 0 && shape[i] != -1) shape_pdt *= shape[i];
+        for (size_t i = 0; i < in_dims_vec.size(); ++i)
+          if (in_dims_vec[i] != 0 && in_dims_vec[i] != -1)
+            in_dims_pdt *= in_dims_vec[i];
+        PADDLE_ENFORCE_EQ(
+            shape_pdt,
+            in_dims_pdt,
+            common::errors::InvalidArgument(
+                "Provided sizes don't multiply up to the size of dim given "
+                "in the input tensor"));
+      }
+    }
    PADDLE_ENFORCE_EQ(unk_dim_idx,


代码逻辑没问题。请添加些注释说明下当前逻辑，避免后期维护困难。另外，符号推导中是否需要同步修改？ValidateShape应该是infermeta中的辅助函数，请检查符号推导中是否存在同样的辅助函数以及unflatten的符号推导是否需要修改

已进行相关修改

DanielSun11 · 2025-07-22T07:02:31Z

test/legacy_test/test_unflatten.py

+class TestUnflattenInputZeroSize(TestUnflattenAPI):
+    def set_args(self):
+        self.x = np.random.rand(4, 0, 16).astype('int16')
+        self.axis = 0
+        self.shape = (2, 2)
+        self.shape_is_tensor = False
+


单测不足以覆盖infermeta中新增的逻辑，当前单测只能覆盖 shape_zero_cnt == in_dims_zero_cnt 的情况，请尝试添加shape_zero_cnt > in_dims_zero_cnt的单测，shape_zero_cnt > in_dims_zero_cnt时应该会抛出异常，请参考单测中测错误case的写法。

已进行相关修改

luyl975 · 2025-07-25T05:21:47Z

/re-run all-failed

luyl975 · 2025-07-25T16:47:29Z

/re-run all-failed

luyl975 · 2025-07-26T10:21:10Z

/re-run all-failed

luyl975 · 2025-07-26T12:39:00Z

/re-run all-failed

DanielSun11 · 2025-07-28T02:54:03Z

请merge下最新的develop分支。相差太多了，ci一直报错

luyl975 · 2025-07-28T06:17:01Z

已进行合并

请merge下最新的develop分支。相差太多了，ci一直报错

DanielSun11

LGTM

cangtianhuang

LGTM

luyl975 added 2 commits July 11, 2025 10:35

try to fix 0-size problem about unflatten api

218e51c

add the unit test of 0-size problem about unflatten api

318f867

luotao1 assigned luotao1 and DanielSun11 Jul 11, 2025

luotao1 added the HappyOpenSource Pro 进阶版快乐开源活动，更具挑战性的任务 label Jul 11, 2025

luotao1 mentioned this pull request Jul 11, 2025

【开源任务】 Paddle API 0-size 机制建设 #72637

Closed

paddle-bot bot added the contributor External developers label Jul 11, 2025

add pre-commit

5395b49

luyl975 added 2 commits July 17, 2025 11:02

fix codestyle1

b8e76a0

fix codestyle 2

a1fe5d0

DanielSun11 requested changes Jul 17, 2025

View reviewed changes

.venv/pyvenv.cfg Outdated Show resolved Hide resolved

delete .venv

9b73a83

cangtianhuang suggested changes Jul 20, 2025

View reviewed changes

paddle/phi/infermeta/unary.cc Outdated Show resolved Hide resolved

paddle/phi/infermeta/unary.cc Outdated Show resolved Hide resolved

use paddle_enforce_eq instead a compare by if

4876fe7

DanielSun11 reviewed Jul 22, 2025

View reviewed changes

add unit test

84954ea

new error unit test

65d78eb

new error unit test2

e76b3de

test

536b210

luotao1 changed the title ~~[0-size Tensor No. 353] Add 0-size Tensor support for unflatten API.~~ [0-size Tensor No.353] Add 0-size Tensor support for unflatten API. Jul 26, 2025

add 0-size unit test 2

9761026

luyl975 added 2 commits July 28, 2025 14:11

change unit

6b65f6d

Merge remote-tracking branch 'upstream/develop' into feature/unflatten

26e14ed

DanielSun11 approved these changes Jul 28, 2025

View reviewed changes

cangtianhuang approved these changes Jul 28, 2025

View reviewed changes

DanielSun11 merged commit 77903fa into PaddlePaddle:develop Jul 28, 2025
70 of 71 checks passed

Conversation

luyl975 commented Jul 11, 2025 • edited by DanielSun11 Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Uh oh!

DanielSun11 commented Jul 15, 2025

Uh oh!

luyl975 commented Jul 16, 2025

Uh oh!

DanielSun11 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

codecov-commenter commented Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

luyl975 commented Jul 18, 2025

Uh oh!

Uh oh!

Uh oh!

luyl975 commented Jul 21, 2025

Uh oh!

luyl975 commented Jul 21, 2025

Uh oh!

luyl975 commented Jul 22, 2025

Uh oh!

DanielSun11 Jul 22, 2025

Choose a reason for hiding this comment

Uh oh!

luyl975 Jul 24, 2025

Choose a reason for hiding this comment

Uh oh!

DanielSun11 Jul 22, 2025

Choose a reason for hiding this comment

Uh oh!

luyl975 Jul 24, 2025

Choose a reason for hiding this comment

Uh oh!

luyl975 commented Jul 25, 2025

Uh oh!

luyl975 commented Jul 25, 2025

Uh oh!

luyl975 commented Jul 26, 2025

Uh oh!

luyl975 commented Jul 26, 2025

Uh oh!

DanielSun11 commented Jul 28, 2025

Uh oh!

luyl975 commented Jul 28, 2025

Uh oh!

DanielSun11 left a comment

Choose a reason for hiding this comment

Uh oh!

cangtianhuang left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

luyl975 commented Jul 11, 2025 •

edited by DanielSun11

Loading

codecov-commenter commented Jul 17, 2025 •

edited

Loading