Inconsistant results between numpy.median() and torch.median()

[lijunzh]

Numpy seems to give the correct median of even number of elements (which is the mean of the center two elements) as defined in Wikipedia while torch gives one of the center elements that are closer to their mean (I guessed from the results I see). I am not sure if this is intended behavior or we need to fix this bug? As least, I think we should have a function that does the standard median for comparision with numpy or any other math programs, such as MATLAB.

In [88]: a = np.random.randn(4, 4)

In [89]: a
Out[89]:
array([[ 0.21654775, -2.84948564, -0.89086479,  0.31074037],
       [-2.02333919,  0.59465567,  1.55680421, -0.33646373],
       [ 1.15586001,  0.16046311,  0.01646207, -1.19663499],
       [ 0.7947269 , -0.22558656, -1.25525967, -0.29217645]])

In [90]: b = torch.from_numpy(a)

In [91]: np.median(a, axis=0)
Out[91]: array([ 0.50563732, -0.03256172, -0.43720136, -0.31432009])

In [92]: torch.median(b, dim=0)[0]
Out[92]:

 0.2165 -0.2256 -0.8909 -0.3365
[torch.DoubleTensor of size 1x4]

[lijunzh]

By the way, it seems that PyTorch's tensor in cuda does not have a median implemenation yet. I got this error message when trying to call it:

AttributeError: 'torch.cuda.DoubleTensor' object has no attribute 'median'

or

TypeError: Type torch.cuda.DoubleTensor doesn't implement stateless method median

Maybe it should be a separate issue, but I think I should mention it here since it is also related to median funciton.

[soumith]

wrt current behavior, it is intended and we will not be fixing it.

wrt median being defined for cuda, it is now defined on master and will be present in the next release.

[fmassa]

@lijunzh for some background on the decision to implement median as is, see torch/torch7#182

[stiv-yakovenko]

HI! I've learned how to emulate np.median with torch:

import torch
import numpy as np
y =[1, 2, 3, 5, 9, 1]
print("numpy=",np.median(y))
print(sorted([1, 2, 3, 5, 9, 1]))
yt = torch.tensor(y,dtype=torch.float32)
ymax = torch.tensor([yt.max()])
print("torch=",yt.median())
print("torch_fixed=",(torch.cat((yt,ymax)).median()+yt.median())/2.)

[lijunzh]

@stiv-yakovenko Thanks for the suggestion.

ymax = torch.tensor([yt.max()])

This line can be simplified as ymax = yt.max()[None].

[sztal]

Hi, I know this issue is closed already, but I wonder why PyTorch has to implement median in such a inconsistent way with respect to other standard numerical libraries? I just got hit by this weird behavior and it took me a while to figure out what is going on.

I understand that there are some good reasons for this design decision, but I just wonder what they are as the documentation seems not to discuss this issue.

[fxmarty]

It should be stated explicitly in the documentation that torch and numpy median behave differently.