The CLI and example models use a JSON exchange format via
model_io::import_json. The ONNX protobuf foundation (import_onnx) already
validates opset and operators; parameter extraction is tracked in issues #5 /
#6. See onnx-import.md.
The JSON schema mirrors the in-memory model types.
{ "kind": "logistic_regression", "weights": [0.4, -1.2, 0.8], "bias": 0.1 }{
"kind": "decision_tree",
"num_features": 2,
"nodes": [
{ "type": "split", "feature_index": 0, "threshold": 0.5, "left": 1, "right": 2 },
{ "type": "leaf", "value": 0.0 },
{ "type": "leaf", "value": 1.0 }
]
}{
"kind": "tiny_mlp",
"layers": [
{ "weights": [1.0], "biases": [0.0], "input_size": 1, "output_size": 1 }
]
}All floating-point values are quantized to Q16.16 on import.