ignition-monorepo / backend-onnx/src / ExportResult

Interface: ExportResult

Defined in: backend-onnx/src/exporter.ts:40

Saves a TF.js LayersModel to disk in the format expected by the Python tf2onnx converter, and returns the shell commands to complete the conversion to .onnx.

Full conversion flow

Step 1 (JS) — Call this function to save the model:


import { DQNAgent } from '@ignitionai/backend-tfjs';
import { saveForOnnxExport } from '@ignitionai/backend-onnx';
 
const { modelDir, conversionScript } = await saveForOnnxExport(agent.model, './exports/my_model');
console.log(conversionScript); // Print or execute the Python script

Step 2 (Python) — Run the generated script:


pip install tensorflowjs tf2onnx
tensorflowjs_converter --input_format=tfjs_layers_model \
  ./exports/my_model/model.json ./exports/my_model_savedmodel/
python -m tf2onnx.convert \
  --saved-model ./exports/my_model_savedmodel \
  --output ./exports/my_model.onnx \
  --opset ${opset}

Step 3 (JS) — Load with OnnxAgent:


const agent = new OnnxAgent({ modelPath: './exports/my_model.onnx', actionSize: 4 });
await agent.load();

Why Python?

No maintained npm package performs reliable TF.js → ONNX conversion. The official path is: TF.js JSON → TF SavedModel (via tensorflowjs_converter) → ONNX (via tf2onnx).

Properties

modelDir

modelDir: string

Defined in: backend-onnx/src/exporter.ts:42

Directory where the TF.js model was saved (model.json + weights.bin)

conversionScript

conversionScript: string

Defined in: backend-onnx/src/exporter.ts:44

Shell script (bash) that converts the saved model to .onnx