Incorrect number of active fields for NeuralNetwork PMML about jpmml-evaluator HOT 3 CLOSED

Your PMML looks fine in this regard.

The important thing to understand is that Evaluator#getActiveFields() defines the "public API" of your NN model, whereas NeuralInput elements define the first layer of the "private API" of your NN model.

For example, take the "TxnChannelCodeATM" field. The first element of Evaluator#getActiveFields() is an org.jpmml.evaluator.InputField object, that has name=TxnChannelCodeATM, opType=categorical and dataType=string; there are further Evaluator API methods that let you query its valid value space, which would be returned as a set of ATM, POS and OnL. If you execute your NN model then you would specify the corresponding input value like this:

Evaluator evaluator = ..
Map<FieldName, Object> arguments = ...
InputField tccaInput = (evaluator.getInputFields()).get(0); // In real life, use field name-based lookup, not positional lookup
arguments.put(tccaInput.getName(), tccaInput.prepare("POS"));
Map<FieldName, ?> result = evaluator.evaluate(arguments);

The NeuralNetwork element takes the input value from "public API" layer and can do whatever it likes with it in "private API" layer. Typically, categorical input values are one-hot-encoded, which means that the "TxnChannelCodeATM" would be expanded into three binary indicator fields TxnChannelCodeATM==ATM, TxnChannelCodeATM==POS and TxnChannelCodeATM==OnL (these are your first three NeuralInput elements). However, an NN model might choose to ignore some categories, combine several categories into one, compute interactions between the categories of different input fields etc.

Some other notes. You don't need to set various "numberOf" attributes such as DataDictionary@numberOfFields, NeuralInputs@numberOfInputs, NeuralLayer@numberOfNeurons, NeuralOutputs@numberOfOutputs etc. Also, you define 131 NeuralInput elements, but only use 10 of them (referenced via Con@from attributes, on lines 2026 -- 2048). Perhaps your NN model is missing some middle layers?

from jpmml-evaluator.

Thanks @vruusmann. When i debug this code
InputField tccaInput = ((evaluator.getInputFields()).get(0));
I do see values of the categorical field like ATM,POS in values array but i am not able access them. Is that method not exposed??

from jpmml-evaluator.

What exactly do you mean by "not able to access them"? Did you use the code example that is provided in the "Querying the data schema of models" section of README.md file?

Anyway, the lesson here is that you're only supposed to care about model input fields at a "public API" level. If the Evaluator#getInputFields() returns a 100-element list, then this is the information that you're supposed to work with. Internally, a NN model may keep those 100 input values as-is, or expand them to 132 transformed input fields (eg. by binarizing the values of categorical input fields) - it's a low-level algorithmic detail.

from jpmml-evaluator.