【Hackathon 5th No.33】Fix test_atleast_nd at Mac -part

[megemini]

PR types

Bug fixes

PR changes

Others

Description

验证 mac 下单测失败 #58323

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[megemini]

Update 20231119

之前测试用例中存在两个问题：

1. numpy 版本低，抛出不同的异常类型

问题现象

之前出现的测试挂掉：

单测 : test_atleast_nd, PR: 57785,CI : PR-CI-Mac-Python3,url : https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/9541163
单测 : test_atleast_nd, PR: 59070,CI : PR-CI-Mac-Python3,url : https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/9541296
单测 : test_atleast_nd, PR: 57879,CI : PR-CI-Mac-Python3,url : https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/9541705
单测 : test_atleast_nd, PR: 59040,CI : PR-CI-Mac-Python3,url : https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/9541888

问题分析

分析日志可以看到：

2023-11-16 20:49:54 Requirement already satisfied: numpy>=1.20 in /Users/paddle/Library/Python/3.9/lib/python/site-packages (from -r python/unittest_py/requirements.txt (line 14)) (1.23.5)
2023-11-16 20:49:55 Collecting numpy>=1.20 (from -r python/unittest_py/requirements.txt (line 14))
2023-11-16 20:49:55   Obtaining dependency information for numpy>=1.20 from https://files.pythonhosted.org/packages/b1/c0/563ef35266a30adfb9801bd1b366bc4f67ff9cfed5e707ae2831b3f6a27c/numpy-1.26.2-cp39-cp39-macosx_10_9_x86_64.whl.metadata
2023-11-16 20:49:55   Using cached numpy-1.26.2-cp39-cp39-macosx_10_9_x86_64.whl.metadata (61 kB)

环境中的 numpy 版本为 1.23.5，注意，后面的 1.26.2 是分析依赖使用，不是安装的版本。

此时，对于单测中的混合收入，如：(123, [123]) 只会提示遗弃警告，而不是抛出异常：

In [1]: import numpy as np

In [2]: np.array((123, [123]))
<ipython-input-2-ed5359ff8a8d>:1: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray.
  np.array((123, [123]))
Out[2]: array([123, list([123])], dtype=object)

In [3]: np.__version__
Out[3]: '1.23.5'

由此，/home/shun/Documents/Projects/paddle/megemini/Paddle/python/paddle/base/executor.py 对于混合输入没有抛出 ValueError ，而是日志中的 TypeError：

def _as_lodtensor(data, place, dtype=None):
...
        elif isinstance(data, (list, tuple)):
            data = np.array(data) # 只有 warning，没有 error
            if data.dtype == np.object_:
                raise TypeError(
                    "\n\tFaild to convert input data to a regular ndarray :\n\t* Usually "
                    "this means the input data contains nested lists with different lengths. "
                    "Please consider using 'base.create_lod_tensor' to convert it to a LoD-Tensor."
                )
            data = data.astype(dtype)
        else:
            raise TypeError(
                f"Convert data of type {type(data)} to Tensor is not supported"
            )

对于没有挂掉的 CI，分析日志：

2023-11-18 18:56:06 Requirement already satisfied: numpy>=1.20 in /Users/paddle/Library/Python/3.9/lib/python/site-packages (from -r python/unittest_py/requirements.txt (line 14)) (1.24.4)
2023-11-18 18:56:07 Collecting numpy>=1.20 (from -r python/unittest_py/requirements.txt (line 14))
2023-11-18 18:56:07   Obtaining dependency information for numpy>=1.20 from https://files.pythonhosted.org/packages/b1/c0/563ef35266a30adfb9801bd1b366bc4f67ff9cfed5e707ae2831b3f6a27c/numpy-1.26.2-cp39-cp39-macosx_10_9_x86_64.whl.metadata
2023-11-18 18:56:07   Using cached numpy-1.26.2-cp39-cp39-macosx_10_9_x86_64.whl.metadata (61 kB)

使用 numpy 版本为 1.24.4：

In [1]: import numpy as np

In [2]: np.array((123, [123]))
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[2], line 1
----> 1 np.array((123, [123]))

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2,) + inhomogeneous part.

In [3]: np.__version__
Out[3]: '1.24.4'

此处抛出 ValueError，符合原来单测中 with self.assertRaises(ValueError)，因此可以通过 CI 测试。

解决方案

通过在单测中捕获两种异常解决：

class TestAtleastErrorCombineInputs(BaseTest):
    """test combine inputs, like: `at_leastNd((x, y))`, where paddle treats like numpy"""

    def test_all(self):
        with self.assertRaises((ValueError, TypeError)):
            self._test_dygraph_api(
                self.inputs, self.dtypes, self.shapes, self.names
            )

        with self.assertRaises((ValueError, TypeError)):
            self._test_static_api(
                self.inputs, self.dtypes, self.shapes, self.names
            )

此处没有通过判断 numpy 版本进行分流测试，因为不确定后续 numpy 是否还会变更。

结论

修改前：

• 在 aistudio 环境中更改 numpy 版本为 1.23.5 可以复现错误
• 在 aistudio 环境中更改 numpy 版本为 1.24.4 错误消失

修改后：

• 在 aistudio 环境中更改 numpy 版本为 1.23.5 错误消失，warning 仍然存在
• 在 aistudio 环境中更改 numpy 版本为 1.24.4 错误消失

2. 没有保证在动态图中测试

问题现象

如果在执行单测 TestAtleastAsTensorMethod 前，环境没有进入 动态图，则会抛出异常：

data = var assign_0.tmp_0 : LOD_TENSOR.shape().dtype(int64).stop_gradient(True), dtype = None, place = Place(cpu)
stop_gradient = True

    def _to_tensor_non_static(data, dtype=None, place=None, stop_gradient=True):
    ...
            else:
>               raise TypeError(
                    "Can't constructs a 'paddle.Tensor' with data type {}, data type must be scalar|list|tuple|np.ndarray|paddle.Tensor".format(
                        type(data)
                    )
                )
E               TypeError: Can't constructs a 'paddle.Tensor' with data type <class 'paddle.base.framework.Variable'>, data type must be scalar|list|tuple|np.ndarray|paddle.Tensor

/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/paddle/tensor/creation.py:607: TypeError

问题分析

• 单测执行前应该先保证测试环境，静态图/动态图
• 由于这个测试用例是后来加的，大意了 🫣 ... ...

In [1]: import paddle

In [2]: paddle.enable_static()

In [3]: tensor = paddle.to_tensor(123)

In [4]: paddle.disable_static()

In [5]: tensor.atleast_1d()
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[5], line 1
----> 1 tensor.atleast_1d()

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/paddle/tensor/manipulation.py:4060, in atleast_1d(name, *inputs)
   4058 out = []
   4059 for tensor in inputs:
-> 4060     tensor = paddle.to_tensor(tensor)
   4061     if tensor.dim() == 0:
   4062         result = tensor.reshape((1,))

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/paddle/tensor/creation.py:794, in to_tensor(data, dtype, place, stop_gradient)
    792     place = _current_expected_place_()
    793 if in_dynamic_mode():
--> 794     return _to_tensor_non_static(data, dtype, place, stop_gradient)
    796 # call assign for static graph
    797 else:
    798     re_exp = re.compile(r'[(](.+?)[)]', re.S)

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/paddle/tensor/creation.py:607, in _to_tensor_non_static(data, dtype, place, stop_gradient)
    605     return data
    606 else:
--> 607     raise TypeError(
    608         "Can't constructs a 'paddle.Tensor' with data type {}, data type must be scalar|list|tuple|np.ndarray|paddle.Tensor".format(
    609             type(data)
    610         )
    611     )
    612 if not dtype:
    613     if data.dtype in [
    614         'float16',
    615         'float32',
   (...)
    618         'complex128',
    619     ]:

TypeError: Can't constructs a 'paddle.Tensor' with data type <class 'paddle.base.framework.Variable'>, data type must be scalar|list|tuple|np.ndarray|paddle.Tensor

解决方案

将 tensor 的创建放置到切换为动态图之后：

class TestAtleastAsTensorMethod(unittest.TestCase):
    def test_as_tensor_method(self):
        input = 123

        for place in PLACES:
            paddle.disable_static(place)

            tensor = paddle.to_tensor(input)

可以验证：

In [1]: import paddle

In [2]: paddle.disable_static()

In [3]: tensor = paddle.to_tensor(123)
tens
In [4]: tensor.atleast_1d()
W1119 12:58:47.352586 10509 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 12.0, Runtime API Version: 11.6
W1119 12:58:47.363787 10509 gpu_resources.cc:164] device: 0, cuDNN Version: 8.4.
Out[4]: 
Tensor(shape=[1], dtype=int64, place=Place(gpu:0), stop_gradient=True,
       [123])

结论

修改前：

• 单测 TestAtleastAsTensorMethod 会抛出 TypError

修改后：

• 单测 TestAtleastAsTensorMethod 会无异常

总结：

• atleast_xd 这几个 api 本身应该没什么问题
• 对 numpy 版本不熟悉导致第一个问题出现
• 而第二个问题纯属自己大意了，需要面壁思过 ... ...

另外，单测增加了 set_device 语句，手动切换（动态图）运行设备 ～

@luotao1 @SigureMo