mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Haibin Lin <haibin.lin....@gmail.com>
Subject Re: Duplication of Operators for sampling from random distributions
Date Tue, 24 Jul 2018 21:31:48 GMT
Hi Anirudh,

Thanks for asking this on dev@. I looked at the doc for sample_uniform and
random_uniform, and found that the API is different. For sample_uniform,
the type of arguments `low` and `high` is NDArray, while that of
random_uniform's is float. I don't think they're going to be deprecated.

The recommended API to generate a random number is via the ndarray.random.*
or symbol.random.*, which accept both float and NDArray, and under the hood
invoke either sample_xxx or random_xxx correspondingly.

Best,
Haibin

On Mon, Jul 23, 2018 at 1:42 PM, Anirudh Acharya <anirudhkrec@gmail.com>
wrote:

> Hi All,
>
> I had earlier filed an issue with functionality-duplication/code-refactor
> here - https://github.com/apache/incubator-mxnet/issues/11811
>
> As per the suggestion in the github issue I would like to bring it to the
> attention of the wider community -
>
> The operators defined in sample_op.cc and multisample_op.cc are seemingly
> performing the same tasks. Both these files define the following operators
> respectively
>
> sample_op.cc
> -----------
> random_uniform
> random_normal
> random_gamma
> random_exponential
> random_poisson
> random_negative_binomial
> random_generalized_negative_binomial
>
> multisample_op.cc
> ----------
> sample_uniform
> sample_normal
> sample_gamma
> sample_exponential
> sample_poisson
> sample_negative_binomial
> sample_generalized_negative_binomial
>
> The only difference that I can glean from the documentation is that
> operators in multisample_op.ccperforms concurrent sampling from multiple
> distributions, but the behavior of the operators is not different.
>
> Is sample_op.cc being retained for legacy reasons or backward
> compatibility? Can it be deprecated or EOLed? Correct me if I am wrong
> here.
>
>
> Thanks
>
> Anirudh
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message