Index of evaluateAll output DF does not match that of input about jpmml-evaluator-python HOT 3 CLOSED

Noted.

This fix would make a metadata-level connection between these two DataFrame objects:
https://github.com/jpmml/jpmml-evaluator-python/blob/0.10.1/jpmml_evaluator/__init__.py#L164
https://github.com/jpmml/jpmml-evaluator-python/blob/0.10.1/jpmml_evaluator/__init__.py#L175

Most ML models perform a 1-to-1 mapping (ie. one row of input produces one row of output). But it wouldn't hurt to have a quick check about this assumption, and fall back to a default numbering of output rows if it's being violated.

from jpmml-evaluator-python.

Thanks. Shouldn't there be a meta-level connection between these two objects? i.e.

evaluateAll(df) shouldBeSameAs df.apply(lambda row: evaluate(row), axis=1) #probably with result_type = "expand"

This connection would break if df can be split into one or more groups and the predictions for each group are based on that group's data as a whole. But this (and others such) seems like entirely different use-case that probably deserves its own function.

from jpmml-evaluator-python.

evaluateAll(df) shouldBeSameAs df.apply(lambda row: evaluate(row), axis=1)

The evaluateAll(df) method call involves exactly one Python-Java-Python roundtrip, because all data is sent over in one go (as a dict data structure). If you replace it with elementary evaluate(X) method calls, then there will be many Python-Java-Python roundtrips (one for each data frame row).

It is my guesstimate that the computational cost of Python-Java-Python rountrip could be comparable to the cost of the actual model evaluation work. So, by limiting the number of roundtrips, I'm expecting to get more "useful" work done per unit of time.

For reference, here's the mechanics of a Python-Java-Python roundtrip:

1. In Python, dump arguments in pickle data format to a pickle blob (byte array).
2. Send the pickle blob from Python to Java.
3. In Java, load arguments from pickle blob.
4. In Java, perform model evaluation.
5. In Java, dump results in pickle data format to a pickle blob.
6. Send the pickle blob from Java to Python.
7. In Python, load results from pickle blob.

from jpmml-evaluator-python.