If p = 2 , 4 or 6, the result of
p = 3
my_tensor.pow(p).pow(1/p)
is ok
if p=3, 5 and …
the result with nan values, where negative numbers?
Hi,
This is because any power 1/p is only defined for positive numbers.
But negative number at the power 3,5, etc gives negative numbers.
So you get nan.
PS: You should use backtick ` and not " to format code in a nicer way. I edited you post to make it look better.
Hello Mikhail (and Alban)!
To give a little more context to Alban’s answer:
Yes, pytorch doesn’t want you to take a root of a negative
number (e.g. sqrt (-1) which is (-1)**(1/2), a fractional
power), so it gives you a nan.
(I am using x**y to mean "raise x to the power y".)
When you raise a negative number to an even power you get
a positive number, so there is no problem taking the root. But
when you raise a negative number to an odd power, you get a
negative number, so pytorch gives you nan for the root.
However, if you use complex numbers (numbers that have a
real and so-called imaginary part), roots of negative numbers
make perfect sense. But, for example, there is no square root
of -1 that is purely real (that is, that has no imaginary part).
So, if you stick to real numbers, the best you can do is nan.
Nonetheless, you could argue that the cube root of -1 (that
is, (-1)**(1/3)) is perfectly well defined as the purely real
number -1. Indeed it is (because (-1)**3 == -1). So why
don’t we do this?
Now the fun begins:
A number (including a negative number) has k k-th roots. That
is x**(1/k) has k legitimate values, in that there are k distinct
numbers, z, (most or all of which will be complex numbers) for
which z**k == x.
To make the discussion a little simpler, let’s just look at -1.
If we want to use just one value for (-1)**(1/k) (rather than
k different values), which value should we use? Mathematicians
(for reasons that aren’t chiselled in stone but do make sense)
prefer to use (of the k values) the value of (-1)**(1/k)
that is closest (in the complex plane) to 1. This particular
choice will always be complex (i.e., not purely real), so, if
you want to stick to real numbers, the best you can do is nan.
(You could have chosen -1 – that would be perfectly legitimate,
if less standard, but that’s not what pytorch does*.)
Just for fun, let’s compare pytorch’s Tensor.pow(), python’s **
operator, and python’s math.pow():
import torch
print (torch.__version__)
t = torch.Tensor ([1, -1])
for p in [2, 3, 4, 5]:
print ('p = ', p)
print (' pytorch-powpow = ', t.pow (p).pow (1 / p))
import math
for p in [2, 3, 4, 5]:
print ('p = ', p)
try:
print (' math-powpow = ', math.pow (math.pow (-1, p), 1/p))
except Exception as e:
print ('Exception = ', e)
for p in [2, 3, 4, 5]:
print ('p = ', p)
print (' **-powpow = ', ((-1)**p)**(1/p))
Here is the output of running this script:
0.3.0b0+591e73e
p = 2
pytorch-powpow =
1
1
[torch.FloatTensor of size 2]
p = 3
pytorch-powpow =
1
nan
[torch.FloatTensor of size 2]
p = 4
pytorch-powpow =
1
1
[torch.FloatTensor of size 2]
p = 5
pytorch-powpow =
1
nan
[torch.FloatTensor of size 2]
p = 2
math-powpow = 1.0
p = 3
Exception = math domain error
p = 4
math-powpow = 1.0
p = 5
Exception = math domain error
p = 2
**-powpow = 1.0
p = 3
**-powpow = (0.5000000000000001+0.8660254037844386j)
p = 4
**-powpow = 1.0
p = 5
**-powpow = (0.8090169943749475+0.5877852522924731j)
For an odd root of a negative number, pytorch and python’s
math.pow() basically agree. (Pytorch gives nan, and python
throws an exception.)
But python’s ** operator give you a complex number (and,
indeed, of the k roots, the complex number that is closet to
1), and (for the case of (-1)**(1/k)) chooses not to give
you -1.
*) In order for pytorch to be smart enough to give you -1 for,
say, torch.Tensor ([-1]).pow (1/3), it would have to know
that the argument to pow() is, in fact one-third. But all pytorch
sees for this argument is a floating-point number that is very
close to, but not exactly one-third. So figuring out that you meant
precisely one-third would be quite tricky.
Have fun!
K. Frank
2 Likes
Thanks a lot, now I understand why it happened.
I decided to do as in LPPool1d class:
p = 3
torch.sign(my_tensor) * torch.abs(my_tensor.pow(p)).pow(1. / p)
maybe it will be useful for someone