Hi Matthias,
thanks for taking a look at the document!
Let me try to answer your questions with some ideas - part of this POC
and my current work is to find out what the best answers are!
1) I see basically two usecases for this DSL:
- users write functions/algorithms much like prepared statements in
SQL (defining functions `def fun(a: T, b: U) = parallelize { ... }` and
executing them later)
- users interactively submit snippets to SystemML (using `val A =
parallelize { C %*% D } execute()` and directly executing)
In general, we should probably offer a write() primitive like in DML
that persists the data on the filesystem. In the second case it's not
quite clear to me what would be the best option right now. Intuitively I
would want the result to be of the same type that my initial DSL
expression was. If I multiply two matrices for example, I would want a
Matrix (DSL Type) as a result. Ideally, I would not have to care about
what underlying representation the actual matrix has and could just use
the result in my next statement/function until I would want to pass the
result somewhere else (persist it, transform it into a spark dataframe
etc.). Given that right now the Algorithm.execute() would take the
generated DML string and execute it using the MLContext, we would be
free to return anything that the context can return - or wrap it in the
DSL Matrix type. I am happy to discuss what would be best here!
For reusing the MLContext, I suggest using a global context that is held
via a lazy variable in the api package object that is imoprted when
using the DSL. The run method would get an implicit argument of type
MLContext and the user would not have to take care of passing it. The
laziness will help reusing it.
2) I think it should be possible to formulate semantically equivalent
operations using breeze - the question is if the maintenance and
implementation of two operational APIs makes sense and is feasible. The
breeze rapid prototyping would be very nice IMO but probably shouldn't
become a major source of work. As for the DNN operations - we could
probably find a way of wrapping those, too - but I don't really think it
makes sense and we might think about how we want to offer DML libraries
in our DSLs in general. Apart from that, it seems like it is possible to
call java functions directly from DML - this might be an interesting
aspect to keep in mind for UDFs.
3) A frame datatype should definitely be part of the DSL and would
probably work very similar to the Matrix abstraction. Right now I am
working with matrices to figure out how a good way to use the DSL would
look like. Apart from the general goal and idea of an embedded DSL, this
includes figuring out what is possible in DML (and SystemML in general).
The goal should be a DSL that allows for full support of all DML
features (possibly even more).
I hope this clarifies some of your questions and I will send updates on
the progress and update the document as I go.
Thanks!
Felix
Am 24.09.2016 10:11 schrieb Matthias Boehm:
thanks for sharing the summary - this is very nice. While looking over
the example, I had the following questions:
1) Output handling: It would be great to see an example how the
results of Algorithm.execute() are consumed. Do you intend to hand out
our binary matrix representation or MLContext's Matrix from which the
user then requests specific output formats? Also if there are multiple
Algorithm instances, how is the MLContext (with its internal state of
lazily evaluated intermediates) reused?
2) Scala-breeze prototyping: How do you intend to support operations
that are not supported in breeze? Examples are removeEmpty, table,
aggregate, rowIndexMax, quantile/centralmoment, cummin/cummax, and DNN
operations?
3) Frame data type and operations: Do you also intend to add a frame
type and its operations? I think for this initial prototype it is not
necessarily required but please make the scope explicit.
Regards,
Matthias
fschueler---09/23/2016 04:36:14 PM---As discussed in the related Jira
(SYSTEMML-451) I have started to implement a prototype/proof of co
From: fschue...@posteo.de
To: dev@systemml.incubator.apache.org
Date: 09/23/2016 04:36 PM
Subject: Proof of Concept: Embedded Scala DSL
-------------------------
As discussed in the related Jira (SYSTEMML-451) I have started to
implement a prototype/proof of concept for an embedded DSL in Scala.
I have summarized the current approach in a short document that you
can
find on github together with the code:
https://github.com/fschueler/emma/blob/sysml-dsl/emma-sysml-dsl/README.md
[1]
Please note that current development happens in the Emma project but
will move to an independent module in the SystemML project once the
necessary additions to Emma are merged. By having the DSL in a
separate
module, we can include Scala and Emma dependencies only for the users
that actually want to use the Scala DSL.
The current code serves as a proof of concept to discuss further
development with the SystemML community. I especially welcome input
from
SystemML Scala users on the usability of the API design.
Next steps will include the translation from Scala code to DML with
support of all features currently supported in DML, including control
flow structures.
Also, a coherent way of executing the generated scripts from Scala and
the interaction with outside data formats (such as Spark Dataframes)
will be integrated.
I am happy to answer your questions and discuss the described approach
here!
Felix
Links:
------
[1]
https://github.com/fschueler/emma/blob/sysml-dsl/emma-sysml-dsl/README.md
