Sure, let me give the high-level goal and some context. Keras2DML and
Caffe2DML take Keras and Caffe model specifications and automatically
generate DML scripts similar to ./scripts/nn/examples/mnist_lenet.dml that
can be executed by SytemML. These scripts have a common structure of (1)
iterating over epochs and batches per epoch, (2) executing the forward and
backward passes, (3)  calling the optimizer update, and (4) computing the
training/validation accuracy every few batches. This default generation
approach will remain unchanged.

However, in order to optionally leverage the new paramserv builtin function
that SYSTEMML-2083 aims to introduce, a couple of extensions are required.
First, the programmatic APIs of Keras2DML and Caffe2DML need to be extended
to allow specifying the use of paramserv and its parameters. Second, we
need to change the actual script generator, because paramserv handles the
loop over epochs and batches itself. An initial idea is to generate
functions for (2)-(3) above as well as a call to paramserv which passes
these functions as arguments.

Obviously, the task SYSTEMML-2089 is closely related to SYSTEMML-2084
because the builtin function API design (e.g., its signature) directly
defines the required extensions.

Regards,
Matthias

On Sat, Mar 17, 2018 at 11:49 AM, Govinda Malavipathirana <
mp.govi...@gmail.com> wrote:

> Hi
> Still there some cloud around the task. I read the documentation about
> Caffe2DML, Keras2DML. Could I know current scope, expected
> tasks(implementations), need an explanation. Hope we gonna implemnet
> ResNet50 with Caffe2DML like missing parts of the existing script
> generators.
> Regards,
> Govinda.
>
>
> On Fri, Mar 16, 2018 at 4:42 AM Niketan Pansare <npan...@us.ibm.com>
> wrote:
>
>> Hi Govinda,
>>
>> Please see the below responses:
>>
>>    1. As a example keras do support theano, tensorflow backends, in this
>>    script generator are we going to use theano/tensorflow backends as well? 
>> if
>>    it does the script conversion should happen like this(I guess)..We donot
>>    use any backend, we only map Keras API to DML script.
>>
>>    2. How is the mapping functionality happen between framework script
>>    and DML script?
>>    Please refer to design documentation: https://github.com/apache/
>>    systemml/blob/master/src/main/scala/org/apache/sysml/api/dl/
>>    Caffe2DML.scala#L56
>>    
>> <https://github.com/apache/systemml/blob/master/src/main/scala/org/apache/sysml/api/dl/Caffe2DML.scala#L56>
>>
>>    3. And what could be the limitation, such as are we define all the
>>    DML functionality in terms of keras/caffe?
>>    Adding custom python layers
>>
>>
>> Thanks,
>>
>> Niketan Pansare
>> IBM Almaden Research Center
>> E-mail: npansar At us.ibm.com
>> http://researcher.watson.ibm.com/researcher/view.php?person=us-npansar
>>
>> [image: Inactive hide details for Govinda Malavipathirana ---03/15/2018
>> 12:38:42 PM---Hi' I'm currently working on the GSoC proposal an]Govinda
>> Malavipathirana ---03/15/2018 12:38:42 PM---Hi' I'm currently working on
>> the GSoC proposal and looking for contribute on
>>
>> From: Govinda Malavipathirana <mp.govi...@gmail.com>
>> To: Matthias Boehm <mboe...@gmail.com>, dev@systemml.apache.org
>> Date: 03/15/2018 12:38 PM
>> Subject: [SYSTEMML-2089] Extended Caffe2DML and Keras2DML script
>> generators.
>> ------------------------------
>>
>>
>>
>> Hi'
>> I'm currently working on the GSoC proposal and looking for contribute on
>> Caffe2DML, Keras2DML script generators. I would like to know the sub task
>> in details. As far as I can understand when user writes a Caffe, Keras
>> script it should be convert to the DML script. In such case I have a couple
>> of questions,
>>
>>    1. As a example keras do support theano, tensorflow backends, in this
>>    script generator are we going to use theano/tensorflow backends as well? 
>> if
>>    it does the script conversion should happen like this(I guess)...
>>
>>    ​
>>    2. How is the mapping functionality happen between framework script
>>    and DML script?
>>    3. And what could be the limitation, such as are we define all the
>>    DML functionality in terms of keras/caffe?
>>
>>  Also I'm hoping to contribute on [SYSTEMML-2085] Basic runtime
>> primitives as well hope to get the details of it as well.
>> Best regards,
>> Govinda.
>>
>> ​
>>
>>
>>

Reply via email to