API improvement for paddle.median 易用性提升 #62407
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
zhwesky2010
changed the title
| @@ -367,11 +367,18 @@ def median(x, axis=None, keepdim=False, name=None): | |||
| the output Tensor is the same as ``x`` except in the reduced | |||
| dimensions(it is of size 1 in this case). Otherwise, the shape of | |||
| the output Tensor is squeezed in ``axis`` . Default is False. | |||
| mode (str, optional): Whether to use mean or min operation to calculate | |||
python/paddle/tensor/stat.py
Outdated
| name (str, optional): Name for the operation (optional, default is None). | ||
| For more information, please refer to :ref:`api_guide_Name`. | ||
|
|
||
| Returns: | ||
| Tensor, results of median along ``axis`` of ``x``. If data type of ``x`` is float64, data type of results will be float64, otherwise data type will be float32. | ||
| ((Tensor, Tensor), optional), results of median along ``axis`` of ``x``. If ``mode`` is |
There was a problem hiding this comment.
这里语句写的有些绕了，不直白，分情况列举吧：
mode='avg'，返回一个Tensor
mode='min'，指定axis，返回一个Tensor，不指定axis，返回两个Tensor
数据类型再单独列举
python/paddle/tensor/stat.py
Outdated
| ), | ||
| dtype=dtype, | ||
| ) | ||
| if inp_axis is not None: |
There was a problem hiding this comment.
已删除对 out_idx 的 cast ，仅对 out_tensor 做 cast
python/paddle/tensor/stat.py
Outdated
| else: | ||
| out_tensor = paddle.cast( | ||
| paddle.slice( | ||
| tensor_topk, axes=[axis], starts=[kth], ends=[kth + 1] | ||
| ), | ||
| dtype=dtype, | ||
| ) | ||
| if inp_axis is not None: |
python/paddle/tensor/stat.py
Outdated
| @@ -423,6 +442,9 @@ def median(x, axis=None, keepdim=False, name=None): | |||
| ], 'when input 0-D, axis can only be [-1, 0] or default None' | |||
| is_flatten = True | |||
|
|
|||
| if mode not in ('avg', 'min'): | |||
| raise ValueError(f"Mode {mode} is not supported. Must be avg or min.") | |||
| inp_axis = axis | |||
There was a problem hiding this comment.
我想在后面处理的时候用 axis 的原始输入是否为 None 做一个判断，但是 axis 为 None 的话下面那一步会把 axis 的值改为 0, 之后就没办法用 axis 来判断了
if axis is None:
is_flatten = True
if is_flatten:
x = paddle.flatten(x)
axis = 0
python/paddle/tensor/stat.py
Outdated
| tensor_topk, axes=[axis], starts=[kth], ends=[kth + 1] | ||
| ) | ||
| out_tensor = paddle.cast(out_tensor, dtype=dtype) / 2 | ||
| else: # mode == 'min' |
There was a problem hiding this comment.
这个cast其实也不是特别合理，目前就mode == 'avg'维持现状吧，新增的mode='min'就不cast了，torch应该也是这样吧
| @@ -22,7 +23,84 @@ | |||
| DELTA = 1e-6 | |||
|
|
|||
|
|
|||
| class TestMedian(unittest.TestCase): | |||
| def np_medain_min(data, keepdims=False): | |||
There was a problem hiding this comment.
np.median 本身是取平均的，所以在 TestMedianAvg 这个类里面就用 np.median 对齐， 在 TestMedianMin 里面用自己写的 np_medain_min
test/legacy_test/test_median.py
Outdated
| self.assertRaises(ValueError, paddle.median, paddle.to_tensor([])) | ||
|
|
||
|
|
||
| class TestMedianMin(unittest.TestCase): | ||
| def check_numpy_res(self, np1, np2): |
There was a problem hiding this comment.
这个就不用再封函数了，目前的标准测试方式是：np.testing.assert_allclose、np.testing.assert_equal
PR types
New features
PR changes
APIs
Description
mode：默认值 'avg' ，另外可取 'min' 。当所需要计算的 tensor 在axis轴上有偶数个元素时， 'avg' 表示计算结果为中间两个数的算术平均值；'min' 则为二者的最小值。mode = 'min'且axis不为 None 时，返回值为 (median_values, median_indices) ；其他情况下返回值为 median_values 的 tensor 。