【PaddlePaddle Hackathon 2】8、为 Paddle 新增 nanmean API #40472
Conversation
|
Thanks for your contribution! |
|
PR格式检查通过，你的PR将接受Paddle专家以及开源社区的review，请及时关注PR动态。 |
python/paddle/tensor/math.py
Outdated
| @@ -967,6 +967,81 @@ def nansum(x, axis=None, dtype=None, keepdim=False, name=None): | |||
| return sum(tmp_tensor, axis, dtype, keepdim, name) | |||
|
|
|||
|
|
|||
| def nanmean(x,axis=None,keepdim=None,name=None): | |||
There was a problem hiding this comment.
keepdim=False is better？keep same with other API？
There was a problem hiding this comment.
thank you for your suggestion! keep same with other API is the better way. i will update this
| axis = [axis] | ||
| if axis == None: | ||
| return paddle.mean(x[~paddle.isnan(x)], keepdim=keepdim,name=name) | ||
| check_variable_and_dtype(x, 'x/input', |
There was a problem hiding this comment.
just 'x' instead of 'x/input'?
There was a problem hiding this comment.
this is refer to the code of paddle.mean.
'x/input' seems to be used only as a input name when raise a type/dtype error.
| if axis == None: | ||
| return paddle.mean(x[~paddle.isnan(x)], keepdim=keepdim,name=name) | ||
| check_variable_and_dtype(x, 'x/input', | ||
| ['uint16', 'float16', 'float32', 'float64'], |
There was a problem hiding this comment.
dtype should be the same to description of x above. eg: 'uint16' is not in "x (Tensor): The input Tensor with data type float32, float64."
There was a problem hiding this comment.
This is refer to the code of paddle.mean
Although it require the input tensor must with data type of float32/64, but it also check allow data type of 'uint16' and 'float16' code is here.
Because of the issus describe here(paddle.nanmean extends the functionality of the paddle.mean API), so i just follow the code of paddle.mean
python/paddle/tensor/math.py
Outdated
| if axis == None: | ||
| return paddle.mean(x[~paddle.isnan(x)], keepdim=keepdim,name=name) |
There was a problem hiding this comment.
can the logic of code L1041~L1042 below cover this branch? we need to handle the condition 'axis == None' alone?
There was a problem hiding this comment.
thank you for your suggestion!
the logic of code L1041~L1042 below can cover this branch.
we don't need handle this condition alone.
i will take this advise and update this problem.
|
|
||
| def setUp(self): | ||
| self.x_shape = [2, 3, 4, 5] | ||
| self.x = np.random.uniform(-1, 1, self.x_shape).astype(np.float32) |
There was a problem hiding this comment.
self.x does not have 'nan', we should cover all the test case in rfcs
also should include check gradient
There was a problem hiding this comment.
thanks for your suggestion! i will cover all the test case in next update.
but i am confused in check gradient, because i can't find the example of check gradient in the paddle.mean test file test_mean_op.py .
I will appreciate it if you can give me a example.
|
hi~有些细节需要注意一下 @Li-fAngyU |
Li-fAngyU
changed the title
|
@Ligoml Done! |
修改了nanmean的axis参数的文档描述。
|
祝贺你，你的PR测试通过，后续将会纳入飞桨的发版计划中，感谢你对飞桨开发者社区的参与。 |
|
你的PR有最新反馈，请及时修改。 |
updata nanmean's sample code (:name: code-example1)
修改nanmean的example code 错误
update example code
update example code of nanmean
PR types
New features
PR changes
APIs
Describe
ISSUE链接:#40327
RFC的PR链接:PaddlePaddle/community#48
中文文档PR链接:PaddlePaddle/docs#4294
为 Paddle 新增 nanmean API
paddle.nanmean 扩展了 paddle.mean API 的功能，如果输入Tensor中有nan值， paddle.mean在计算中会将涉及nan值的结果都置为nan，而 paddle.nanmean 会跳过nan值。比如输入数据 x = [[nan, 1. , 2. ], [3. , 4. , 5. ]]，x.mean() 得到 [nan]，x.mean(0) 得到 [nan, 2.5, 3.5]，x.nanmean() 得到 [3.]，x.nanmean(0) 得到 [3., 2.5, 3.5]。此API需支持的调用路径为：paddle.nanmean 和 Tensor.nanmean 。