Hyukjin Kwon created SPARK-61873:
------------------------------------

             Summary: Fit, load and copy OneVsRest as PySpark does
                 Key: SPARK-61873
                 URL: https://issues.apache.org/jira/browse/SPARK-61873
             Project: Spark
          Issue Type: Bug
          Components: Connect
    Affects Versions: connect-rust-4.4.0
            Reporter: Hyukjin Kwon


The native OneVsRest of the drop-in and the Rust API differ from PySpark 
4.3.0's in its fit, its loads and its copy:
- The fit runs natively for a subclass of a native classifier, skipping the 
subclass's fit or _fit, where PySpark calls classifier.fit.
- The class count is max(cast(label AS DOUBLE)) + 1 of backticked columns, 
where PySpark's is int(dataset.agg({labelCol: "max"}).head()[...]) + 1 of 
dataset.select(labelCol, featuresCol): string labels such as "10", "9", "0" fit 
11 models instead of 10, a label that int() refuses fails otherwise, and a 
dotted label column fits where PySpark raises UNRESOLVED_COLUMN. The errors of 
a missing classifier, null params and a parallelism below 1 differ too.
- The fit persists the dataset itself, where PySpark's _cache_spark_dataset 
caches it or writes it under SPARKML_TEMP_DFS_PATH.
- The loads refuse the JVM's class names in the metadata, where PySpark's 
readers check no class, and OneVsRestModel.load sends twice the requests of 
PySpark's.
- A parallelism beyond a long is clamped at the save and load, where PySpark 
saves the exact int.
- The Rust OneVsRestModel's apply_param_map does not copy the submodels with 
the map.




--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to