[yocto] Review request 0/13: Contribute meta-tensorflow to Yocto

[Hongxu Jia]

On 2019/2/24 上午1:04, Khem Raj wrote:
> On Sat, Feb 23, 2019 at 7:29 AM Richard Purdie
> <richard.purdie at linuxfoundation.org> wrote:
>> On Fri, 2019-02-22 at 20:49 +0000, Manjukumar Harthikote Matha wrote:
>>>> You might be interested in the yocto layers for tensorflow,
>>>> tensorflow-lite and
>>>> caffe2 on github here [1]. I'm not part of the team that developed
>>>> that work but I
>>>> forwarded your announcement to them. Perhaps there is the
>>>> opportunity for some
>>>> collaboration on the platform independent parts. The maintainer
>>>> details are in the
>>>> readme.
>>>>
>>> Thanks for the layer Hongxu. I agree with Steve, it would be good if
>>> you could collaborate with meta-renesas-ai and introduce the layer as
>>> meta-ai under meta-openembedded.
>> Please don't do the meta-openembedded part!
>>
> I would agree to not make it a sub layer under meta-openembedded, but it can
> be hosted on openembedded git infrastructure, I dont see much problem with that
> if thats the case
>
>> I believe that meta-oe is too large to be maintainable and that we need
>> a larger number of smaller layers.
>>
> There is a fine balance to be had, that I have come to realize over years now
> but AI is large enough and segmented enough to have a layer of its own.
>
>> Having tensorflow in its own layer which as a specific purpose and its
>> specific maintainers who understand it is in my view much more
>> desirable and sustainable.
> I think its a good idea to have various AI infras in one layer
> including tensorflow
> unless we have large enough dev community to maintain each of them so I like
> meta-ai conceptually.

I know to create a standalone meta-ai than meta-tensorflow is more 
reasonable, that's my initial

layer naming, but

- It will dramatically increase the maintainer burden, so I limit the 
scope to the specific framework

   name. There are lots of TODO in tensorflow and I am afraid I do not 
have extra attention to

   other AI framework recently.

- Tensorflow is standalone enough, its build system is google's `bazel', 
like bitbake, it has special

   rules to build everything from scratch. (I've already sent other 
unbazel built recipes to

   meta-openembedded)

- Bazel is built by java, if we do not create sub layer in meta-ai (such 
as meta-ai/meta-tensorflow),

   the number of meta-ai layer deps will be more and more along with 
other AI frameworks

   are added. For other AI framework customer, depends unused layer is 
not a good idea.

- For future AI framework integration, if the framework is huge like 
TensorFlow (another well known is

   Facebook's PyTorch), we could create a standalone layer and appoint 
special maintainer to maintain it;

   if the framework is small and light, or fundamental algorithm 
packages used by multiple frameworks,

   we could create a meta-ai for collection, or directly add them to 
meta-openembedded. (For TensorFlow

   integration, I added 11 fundamental recipes to meta-openembedded )

//Hongxu


>> Cheers,
>>
>> Richard
>>