zhengruifeng created SPARK-23805:
------------------------------------
Summary: support vector-size validation and Inference
Key: SPARK-23805
URL: https://issues.apache.org/jira/browse/SPARK-23805
Project: Spark
Issue Type: Improvement
Components: ML
Affects Versions: 2.4.0
Reporter: zhengruifeng
I think it maybe miningful to unify the usage of \{{AttributeGroup}} and
support vector-size validation and inference in algs.
My thoughts are:
* In \{{transformSchema}}, validate the input vector-size if possible. If the
input vector-size can be obtained from schema, check it.
** Suppose a \{{PCA}} estimator with k=4, the \{{transformSchema}} will
require the vector-size to be no more than 4.
** Suppose a \{{PCAModel}} trained with vectors of length 10, the
\{{transformSchema}} will require the vector-size to be 10.
* In \{{transformSchema}}, inference the output vector-size if possible.
** Suppose a \{{PCA}} estimator with k=4, the \{{transformSchema}} will return
a schema with output vector-size=4.
** Suppose a \{{PCAModel}} trained with k=4, the \{{transformSchema}} will
return a schema with output vector-size=4.
* In \{{transform}}, inference the output vector-size if possible.
* In \{{fit}}, obtain the input vector-size from schema if possible. This can
help eliminating redundant \{{first}} jobs.
Current PR only modifies \{{PCA}} and \{{MaxAbsScaler}} to illustrate my idea.
Since the validation and inference is quite alg-speciafic, we may need to
sperate the task into several small subtasks.
How do you think about this? [~srowen] [~yanboliang] [~WeichenXu123] [~mlnick]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
