ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alexey Zinoviev <zaleslaw....@gmail.com>
Subject Re: ML stable and performance
Date Wed, 18 Sep 2019 09:26:32 GMT
I've started the discussion here
http://apache-ignite-developers.2346864.n4.nabble.com/Proposed-Release-2-7-7-2-7-6-ML-extra-fixed-bugs-td43537.html


пт, 13 сент. 2019 г. в 22:12, Alexey Zinoviev <zaleslaw.sin@gmail.com>:

> The reason was that the last year there is no significant releases of
> Ignite between 2.7 and 2.8, only minor releases with long story of renaming.
> I am and another ML guys are ready in 1-2 months prepare ML module for 2.8
> or for the minor release 2.7.7 = 2.7.6 + updated ML + new fixed bugs
>
> Let's discuss it in separate thread next week
>
>
>
> пт, 13 сент. 2019 г. в 21:55, Denis Magda <dmagda@apache.org>:
>
>> Alexey, I'm wondering,
>>
>> Are there any dependencies on Ignite Core that make us put off the ML
>> changes release until 2.8? I know that you do not support the idea of ML
>> as
>> a separate Ignite module but this concept would allow us to release ML as
>> frequently as we want not being blocked by Ignite core releases.
>>
>>
>> -
>> Denis
>>
>>
>> On Fri, Sep 13, 2019 at 11:45 AM Alexey Zinoviev <zaleslaw.sin@gmail.com>
>> wrote:
>>
>> > I could answer as one of developers of ML module.
>> > Currently is available the ML in version 2.7.5, it supports a lot of
>> > algorithms and could be used in production, but the API is not stable
>> and
>> > will be changed in 2.8
>> >
>> > The ML module will be stable since next release 2.8, also we have no
>> > performance report to compare for example with Spark ML
>> > Based on my exploration the performance of in terms of Big O notation is
>> > the same like in Spark ML (real numbers says that Ignite ML is more
>> faster
>> > than Spark ML due to Ignite in-memory nature and so on)
>> >
>> > Since 2.8 it will have good integration with TensorFlow, Spark ML,
>> XGBoost
>> > via model inference.
>> >
>> > You as a user have no ability to run, for-example scikit-learn or R
>> > packages in distributed mode over Ignite, but you could run the
>> TensorFlow,
>> > using Ignite as a distributed back-end instead of Horovod.
>> >
>> > If you have any questions, please let me know
>> >
>> >
>> >
>> > пт, 13 сент. 2019 г. в 21:28, Denis Magda <dmagda@apache.org>:
>> >
>> >> David,
>> >>
>> >> Let me loop in Ignite dev list that has Ignite ML experts subscribed.
>> >> Please, could you share more details in regards to your performance
>> >> testing
>> >> and objectives for Ignite ML overall?
>> >>
>> >> The module is ready for production and we're ready to help solve any
>> >> cornerstones.
>> >>
>> >> -
>> >> Denis
>> >>
>> >>
>> >> On Fri, Sep 6, 2019 at 4:50 AM David Williams <leeon2013@gmail.com>
>> >> wrote:
>> >>
>> >> > Python is 25 times slower than Java for ML at runtimes, which I found
>> >> out
>> >> > online. But I don't know that statement is true or not. I need
>> insiders'
>> >> > opinion.  Which ml other packages are best options for Ignite?
>> >> >
>> >> > On Fri, Sep 6, 2019 at 7:28 PM Mikael <mikael-aronsson@telia.com>
>> >> wrote:
>> >> >
>> >> >> Hi!
>> >> >>
>> >> >> I have never used it myself but it's been there for long time and
I
>> >> >> would expect it to be stable, and yes it will run distributed,
I
>> can't
>> >> >> say anything about performance as I have never used it.
>> >> >>
>> >> >> You will find a lot of more information at:
>> >> >>
>> >> >> https://apacheignite.readme.io/docs/machine-learning
>> >> >>
>> >> >> Mikael
>> >> >>
>> >> >>
>> >> >> Den 2019-09-06 kl. 11:50, skrev David Williams:
>> >> >> >
>> >> >> >
>> >> >> > I am evaluating ML framework for Java platform. I knew Ignite
has
>> ML
>> >> >> > package.
>> >> >> > But I like to know its stability and performance for production.
>> Can
>> >> >> > Ignite
>> >> >> > ML code run in distribute way?
>> >> >> >
>> >> >> > Except its own ML package, which ml packages are best options
for
>> >> >> Ignite?
>> >> >>
>> >> >
>> >>
>> >
>>
>

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