Hello,

this relates to PR #9982 [MXNET-39]. There, I am implementing two new unary
mathematical functions (log of CDF of standard normal distribution, and its
derivative). As Eric rightly points out, they have clunky names, and while
there are important use cases, they do not lie in the mainstream DNN field
for now. They are, however, certainly part of scipy, and some appear in
TensorFlow as well.

I see Eric's point that specific functions with clunky names should not
pollute the global namespace. Indeed, we already have highly specific
things there (gamma, gammaln).

To solve this issue, I'd propose to open a new namespace (which could be
called 'math', but open for other suggestions). It would join the small
number of other namespaces (random, linalg, ...). In the future, more
specific mathematical functions would go there. gamma, gammaln would stay
in global, but could be replicated in there.

Note that I also wrote (with Eric's help) the recent wrappers for unary,
binary maths operators (before this change, they were wrong for float64),
and the new unit tests for float32, float64, float16. I have some vested
interest, along with other contributors at different CoreML Amazon sites,
in making MXNet ready for non-DNN ML work (mainly Bayesian optimization,
Bayesian ML). My PR for these new functions also includes an improvement of
these unit tests.

​Let me know what you think. And those who have opened up new namespaces,
I'd appreciate some guidance on how to do that.

Bye, Matthias​

-- 
Matthias Seeger

Reply via email to