Features/mean var merge cleanup

[coquelin77]

Description

This is a cleanup of the mean and var functions. The moment merging function is rewritten to be used by both mean and var (it is also possible to expand this function to be used for higher order moments).
The var function also supports multi-dimensional axis arguments (the algorithm is the same as it is for mean)

Fixes: #156

Changes proposed:

• combination of moment merging functions into one function
• add multi-dimensional axis support to var

Type of change

Select relevant options.

• New feature (non-breaking change which adds functionality)
• Documentation update

Are all split configurations tested and accounted for?

• yes
• no

Does this change require a documentation update outside of the changes proposed?

• yes
• no

Does this change modify the behaviour of other functions?

• yes
• no

Are there code practices which require justification?

• yes
• no

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (master@1520c1a). Click here to learn what that means.
The diff coverage is n/a.

@@            Coverage Diff            @@
##             master     #445   +/-   ##
=========================================
  Coverage          ?   96.68%           
=========================================
  Files             ?       57           
  Lines             ?    12134           
  Branches          ?        0           
=========================================
  Hits              ?    11732           
  Misses            ?      402           
  Partials          ?        0

Impacted Files	Coverage Δ
heat/core/tests/test_logical.py	97.57% <ø> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1520c1a...91ef6ed. Read the comment docs.

[mtar]

Some tiny remarks on tinier things I noted

[mtar]

Small inconsistency: When axis is set, integer tensors are not supported any longer. Numpy does support ints and floats, while pytorch only supports floats in general. This behaviour is also in 'mean'.

[coquelin77]

because torch does not support ints so we dont either. if its needed, the easiest way to do this would be to cast the input to a float. I dont think that we need to do this though. i dont think that we need to be so religious to numpy

[mtar]

Would you be so kind to remove the failing test? As of today, torch.logical_xor allows non-bool tensors in the new pytorch version.

[coquelin77]

Would you be so kind to remove the failing test? As of today, torch.logical_xor allows non-bool tensors in the new pytorch version.

yes this is an issue that needs to be addressed. are you working on this? Bjoern wants to put a release out today. I am in a meeting at the moment, but once this is done I will get to work on a fix for the latest version of pytorch

[coquelin77]

Would you be so kind to remove the failing test? As of today, torch.logical_xor allows non-bool tensors in the new pytorch version.

should the fix be to use the bitwise xor?

[mtar]

I will make a pull request

[mtar]

One last thing, I was playing with var and realized that heat uses Bassel's correction as the default, similar to pytorch. numpy,however, uses the biased one as default instead. Shall we keep it that way?

[coquelin77]

One last thing, I was playing with var and realized that heat uses Bassel's correction as the default, similar to pytorch. numpy,however, uses the biased one as default instead. Shall we keep it that way?

i think that Bessel's correction should be the default. But this is mostly personal preference. I think that for general statistical methods (as well as more advanced methods) it will provide more statistically reliable results