Re: pytorch_1.6.0-3_source.changes ACCEPTED into unstable

[Gard Spreemann <gspr@nonempty.org>]

Mo Zhou <lumin@debian.org> writes:

> Hi Gard,
>
> I use anaconda (cuda version of pytorch) for my daily research work.
> I'm curious. Will the CPU version of pytorch satisfy your demand?

Hi,

A lot of what I've been doing lately actually does work nicely with the
CPU version, or at least the early parts of the work does – and then I
offload it to a GPU machine managed by someone else anyway for the
longer computations. So having PyTorch in Debian is already a good step
up for me. And it just feels plainly natural to me that all software I
regularly use is in Debian :-)

> If you use neural networks in single precision, the current version of
> pytorch is supposed to show decent computing performance, since the
> onednn / xnnpack have been enabled already.
>
> As for neural network inference in any integer precision, the current
> version may be a little bit slow. As I have not yet enabled the
> corresponding performance dependencies like FBGEMM and NNPACK, etc.

Thanks for the explanation.

> Besides, I plan to add some autopkgtest scripts that trains a basic
> neural network using bin:dataset-fashion-mnist. This script can not
> only be used as CI testing, but also as a simple "benchmarking" tool.

That's really great news!


 Best,
 Gard