D enters Tiobe top 20
Ola Fosheim Grøstad
ola.fosheim.grostad at gmail.com
Wed Nov 6 10:46:54 UTC 2019
On Wednesday, 6 November 2019 at 09:18:58 UTC, Paolo Invernizzi
wrote:
> Alas with a good deep learning D solution, we will for sure use
> D for that, instead of doing the work in python, and then
> translating the successful one into C++ for performance and
> embedding, for example.
Out of curiosity, does that mean that you prefer to do the
training on your own machines rather than "renting" existing
infrastructure (cloud solutions)? Having little experience with
deep learning (although some with basic ML), I thought the
advantage of using a ready-made like the Python one from Google
is that you have a large set of prewritten libraries from both
Google and other third party contributors that you can compose?
Also, if you create your own training-environment, you would
still have to run it on a cluster with GPUs? Then distill it down
into something that can run well on a single CPU/GPU/SoC?
Maybe a framework can do that translation well... but I guess
then the better option would be to have a dedicated high level
language that translates well to both the training-environment
and the final host environment. That seems more reasonable to me?
Or maybe you talk about the initial preprocessing of data? In
which case you can do it now by interfacing with Python from D?
More information about the Digitalmars-d
mailing list