Skip to Content
API Reference@ignitionai/backend-onnxSrcClassesClass: OnnxAgent

ignition-monorepo


ignition-monorepo / backend-onnx/src / OnnxAgent

Class: OnnxAgent

Defined in: backend-onnx/src/agents/onnx-agent.ts:20 

Inference-only RL agent that loads a pre-trained .onnx model and runs forward passes.

Implements AgentInterface — remember() and train() are no-ops since this agent is meant for production inference, not training.

Usage:

const agent = new OnnxAgent({ modelPath: './model.onnx', actionSize: 4 }); await agent.load(); const action = await agent.getAction([0.1, -0.3, 0.5, 0.2]); agent.dispose();

Implements

Constructors

Constructor

new OnnxAgent(config): OnnxAgent

Defined in: backend-onnx/src/agents/onnx-agent.ts:24 

Parameters

config
modelPath

string | Uint8Array<ArrayBuffer> | Buffer<ArrayBufferLike> = ...

Path to the .onnx model file, or a Buffer/Uint8Array for in-memory usage

actionSize

number = ...

Number of discrete actions the agent can take

executionProviders?

string[] = ...

ONNX execution providers in priority order

inputName?

string = ...

Name of the input tensor in the ONNX graph

outputName?

string = ...

Name of the output tensor in the ONNX graph (Q-values)

Returns

OnnxAgent

Methods

load()

load(): Promise<void>

Defined in: backend-onnx/src/agents/onnx-agent.ts:32 

Loads the ONNX model and creates the inference session. Must be called before getAction().

Returns

Promise<void>

Implementation of

AgentInterface.load


getAction()

getAction(observation): Promise<number>

Defined in: backend-onnx/src/agents/onnx-agent.ts:44 

Runs a forward pass through the ONNX model and returns the greedy action (argmax of Q-values).

Parameters

observation

number[]

State vector as number array

Returns

Promise<number>

Discrete action index in [0, actionSize)

Implementation of

AgentInterface.getAction


remember()

remember(_experience): void

Defined in: backend-onnx/src/agents/onnx-agent.ts:64 

No-op — ONNX agents do not accumulate experience

Parameters

_experience

Experience

Returns

void

Implementation of

AgentInterface.remember


train()

train(): Promise<void>

Defined in: backend-onnx/src/agents/onnx-agent.ts:67 

No-op — ONNX agents do not train

Returns

Promise<void>

Implementation of

AgentInterface.train


inspect()

inspect(): object

Defined in: backend-onnx/src/agents/onnx-agent.ts:73 

Returns input/output tensor names for debugging. Useful when the default inputName/outputName don’t match your ONNX model.

Returns

object

inputs

inputs: string[]

outputs

outputs: string[]


loadFromHub()

loadFromHub(repoId, filename?): Promise<void>

Defined in: backend-onnx/src/agents/onnx-agent.ts:85 

Downloads a .onnx file from HF Hub and loads it as the inference session.

Parameters

repoId

string

HF Hub repo (e.g. “salim4n/my-dqn-model”)

filename?

string = 'model.onnx'

Filename in the repo (default: “model.onnx”)

Returns

Promise<void>


dispose()

dispose(): void

Defined in: backend-onnx/src/agents/onnx-agent.ts:91 

Releases the ONNX inference session

Returns

void

Implementation of

AgentInterface.dispose

Last updated on