Github user jkbradley commented on the pull request:
https://github.com/apache/spark/pull/3637#issuecomment-70022005
@etrain Thanks for the feedback! It would be great to hear thoughts about
what could be made easier. Some questions are:
* Is it easy to implement a new algorithm?
* IMO, LinearRegression shows it can require very little code. But what
non-obvious items make it hard? Some possibilities in my mind are:
* adding setter methods for parameters
* passing type parameters to abstract classes
* implementing copy()
* optimizing transform() to be faster than the default implementation
* Is all of the functionality needed? (E.g., do we not want the various
output columns?)
* I have a hard time thinking of an algorithm which does not have these
concepts, or where it would be difficult to implement them.
@tomerk Thanks for the feedback! Some responses:
* Typed vs. Untyped Estimators: The conversation was moved to the JIRA
(though mainly in the design doc linked from there). Basically, it was decided
not to have strongly typed public interfaces. However, I kept the developer
API (protected) strongly typed interfaces for prediction, where I agree it is
useful. For training/fitting, I removed the typed interface since it is less
useful and also requires passing around more type parameters (FeaturesType,
etc.).
* Wrapper for fit/transform so developers do not have to call
transformSchema and combine ParamMaps: IMO, the current default transform()
implementation suffices. For fit(), I agree it may be good to include; Iâll
add that.
* Parameterize sharedParams mixing to eliminate need for setters: I think
itâs a good suggestion. It would make declaring mixing awkward but would be
shorter overall. @etrain What do you think?
* ScalaReflection: The main issues are with the Java API. Ideally,
developers will be able to write new Estimators/Models in Java. But SQL
reflection works differently there (for now), which is why we specify the
DataType in the first place. Java also has trouble understanding this.type.
I'll push an update soon.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
