Question about input feature mapping

[paranjapeved15]

Ask the question
I am trying to use a onnx model trained in python to run inference on with tribe. I was providing the input to my onnx model in the following format:
float[][] sourceArray = new float[2][7];
for (int i=0; i < 2; i++){
for (int j=0; j < 7; j++) sourceArray[i][j] = 1.0f;
}
OnnxTensor tensor = OnnxTensor.createTensor(env,sourceArray);
Map<String, OnnxTensor> inputs = new HashMap();
inputs.put("input",tensor);

what is the best way in tribuo to do the same? I checked this tutorial (https://tribuo.org/learn/4.3/tutorials/external-models-tribuo-v4.html#Loading-in-an-ONNX-model) but I am not sure how to structure my feature mapping.

Is your question about a specific ML algorithm or approach?
No

Is your question about a specific Tribuo class?
No

System details

• Tribuo version: 4.3.1
• Java version (if appropriate): 17
• OS/Architecture (if appropriate): Mac OS

Additional context
Add any other context or screenshots about the question

[Craigacp]

Tribuo needs to know what the mapping is from the feature names it produces to the feature indices that the ONNX model expects. If your model expects a feature vector (rather than an image or some other kind of structured input) then you need to pass in a DenseTransformer to the ONNXExternalModel, along with the string -> int mapping. The feature names are user controlled in some senses, they either come from the headers/field names of the csv/db/json data, or are constructed as the string representation of padded indices (i.e. 000, 001 etc).

For your example with a feature vector of [2][7] is that a batch size of 2 and a 7 element feature vector?

[paranjapeved15]

Yes @Craigacp, the 2 is a batch. So it really needs a float array with 7 elements as input.

[Craigacp]

Ok, so you'd want to make a Tribuo example containing those 7 features, with appropriate names. Those can be semantic ones if the features actually have names, or just 00,...,06, then supply the mapping as appropriate.

[paranjapeved15]

okay, but what about the key "input" do I not need that in tribuo?

[Craigacp]

It's the last argument when constructing an ONNXExternalModel - https://tribuo.org/learn/4.3/javadoc/org/tribuo/interop/onnx/ONNXExternalModel.html#createOnnxModel(org.tribuo.OutputFactory,java.util.Map,java.util.Map,org.tribuo.interop.onnx.ExampleTransformer,org.tribuo.interop.onnx.OutputTransformer,ai.onnxruntime.OrtSession.SessionOptions,java.nio.file.Path,java.lang.String).

[paranjapeved15]

Oh I see. And at inference time what should be the format of my input? What type of object and what function should I be using to run inference on the Model object? @Craigacp

[Craigacp]

The model exposes a predict function which accepts Iterable<Example<T>> or Example<T>, returning Prediction<T> which contains the predicted values and any confidence scores produced by the model.

[paranjapeved15]

Any code example which shows how to build an Example object from raw values?

[Craigacp]

Not specifically, but you can see how the examples are built in all the DataSource implementations, e.g. this one for loading in LibSVM format data - https://github.com/oracle/tribuo/blob/main/Core/src/main/java/org/tribuo/datasource/LibSVMDataSource.java#L348.