Implementation:Microsoft Onnxruntime OrtSession

Knowledge Sources	Description
Source File	java/src/main/java/ai/onnxruntime/OrtSession.java
Repository	Microsoft/onnxruntime

Domains

• Machine Learning Runtime
• Model Inference
• Session Configuration

Overview

OrtSession wraps an ONNX model and provides inference capabilities. It allows inspection of input/output nodes, scoring input feed dictionaries, accessing model metadata, and profiling. It contains the critical inner classes SessionOptions (for execution provider and optimization configuration), RunOptions (for per-run control), and Result (for managing inference outputs). Sessions are produced by OrtEnvironment.

Description

The OrtSession class implements AutoCloseable and provides:

• Session construction: Three constructors accept model path, byte array, or direct ByteBuffer, each with an OrtAllocator and SessionOptions.
• Input/output inspection: getNumInputs(), getNumOutputs(), getInputNames(), getOutputNames(), getInputInfo(), getOutputInfo() provide model metadata.
• Inference: Multiple run() overloads accept input maps (name to OnnxTensorLike), optional requested output names, pinned outputs, and RunOptions.
• Model metadata: getMetadata() returns OnnxModelMetadata.
• Profiling: getProfilingStartTimeInNs() and endProfiling() control profiling.

SessionOptions inner class provides:

• Optimization levels: NO_OPT, BASIC_OPT, EXTENDED_OPT, LAYOUT_OPT, ALL_OPT
• Execution modes: SEQUENTIAL, PARALLEL
• Threading: setIntraOpNumThreads, setInterOpNumThreads
• Execution providers: addCUDA, addROCM, addCPU, addDnnl, addOpenVINO, addTensorrt, addNnapi, addCoreML, addDirectML, addACL, addArmNN, addXnnpack, addQnn, addWebGPU, addExecutionProvider
• Custom ops: registerCustomOpLibrary, registerCustomOpsUsingFunction
• Initializers: addExternalInitializers, addInitializer
• Configuration: addConfigEntry, setDeterministicCompute, setSymbolicDimensionValue

RunOptions inner class provides logging control, termination, run tagging, config entries, and LoRA adapter activation.

Result inner class wraps inference outputs as an iterable, closeable container with index and name-based access.

Code Reference

Source Location

// File: java/src/main/java/ai/onnxruntime/OrtSession.java
// Package: ai.onnxruntime

Signature

public class OrtSession implements AutoCloseable {
    public long getNumInputs();
    public long getNumOutputs();
    public Set<String> getInputNames();
    public Set<String> getOutputNames();
    public Map<String, NodeInfo> getInputInfo() throws OrtException;
    public Map<String, NodeInfo> getOutputInfo() throws OrtException;
    public Result run(Map<String, ? extends OnnxTensorLike> inputs) throws OrtException;
    public Result run(Map<String, ? extends OnnxTensorLike> inputs, Set<String> requestedOutputs) throws OrtException;
    public Result run(Map<String, ? extends OnnxTensorLike> inputs,
        Set<String> requestedOutputs, Map<String, ? extends OnnxValue> pinnedOutputs,
        RunOptions runOptions) throws OrtException;
    public OnnxModelMetadata getMetadata() throws OrtException;
    public long getProfilingStartTimeInNs() throws OrtException;
    public String endProfiling() throws OrtException;
    public void close() throws OrtException;

    public static class SessionOptions implements AutoCloseable { ... }
    public static class RunOptions implements AutoCloseable { ... }
    public static class Result implements AutoCloseable, Iterable<Map.Entry<String, OnnxValue>> { ... }
}

Import

import ai.onnxruntime.OrtSession;
import ai.onnxruntime.OrtSession.SessionOptions;
import ai.onnxruntime.OrtSession.RunOptions;
import ai.onnxruntime.OrtSession.Result;

I/O Contract

Inputs

Name	Type	Description
inputs	Map<String, ? extends OnnxTensorLike>	Named input tensors for inference
requestedOutputs	Set<String>	Optional subset of output names to compute
pinnedOutputs	Map<String, ? extends OnnxValue>	Pre-allocated output tensors
runOptions	RunOptions	Per-run logging and termination controls

Outputs

Name	Type	Description
Result	OrtSession.Result	Container of named OnnxValues (closeable, iterable)
Result.get(int)	OnnxValue	Output value by index
Result.get(String)	Optional<OnnxValue>	Output value by name

Usage Examples

import ai.onnxruntime.*;
import java.nio.FloatBuffer;
import java.util.*;

OrtEnvironment env = OrtEnvironment.getEnvironment();

try (OrtSession.SessionOptions opts = new OrtSession.SessionOptions()) {
    opts.setOptimizationLevel(OrtSession.SessionOptions.OptLevel.ALL_OPT);
    opts.setIntraOpNumThreads(4);

    // Add CUDA if available
    if (OrtEnvironment.getAvailableProviders().contains(OrtProvider.CUDA)) {
        opts.addCUDA(0);
    }

    try (OrtSession session = env.createSession("/path/to/model.onnx", opts)) {
        // Prepare input
        float[][] inputData = new float[1][784];
        // ... fill inputData ...
        OnnxTensor inputTensor = OnnxTensor.createTensor(env, inputData);

        Map<String, OnnxTensorLike> inputs = Collections.singletonMap("input", inputTensor);

        // Run inference
        try (OrtSession.Result results = session.run(inputs)) {
            float[][] output = (float[][]) results.get(0).getValue();
            System.out.println("Prediction: " + Arrays.toString(output[0]));
        }

        inputTensor.close();
    }
}

Related Pages

• OrtEnvironment.java - Factory for creating sessions
• OnnxTensor.java - Input/output tensor type
• OnnxModelMetadata.java - Model metadata
• NodeInfo.java - Input/output node information
• ai_onnxruntime_OrtSession.c - JNI native implementation
• ai_onnxruntime_OrtSession_SessionOptions.c - SessionOptions JNI implementation