architecture/network/onnx/schema

Leaf JSON wire-format schema for NeatapticTS ONNX-like import and export.

This chapter holds the stable, serializable payload shapes shared by both the exporter and importer: model containers, graph nodes, tensors, metadata records, and explicit Conv/Pool mapping declarations.

Keep this file leaf-only. It should describe the persisted document shape, not runtime execution contexts such as node internals, layer traversal state, or importer/exporter orchestration helpers.

Example:

const model: OnnxModel = {
  graph: {
    inputs: [],
    outputs: [],
    initializer: [],
    node: [],
  },
};

architecture/network/onnx/schema/network.onnx.schema.types.ts

Conv2DMapping

Mapping declaration for treating a fully-connected layer as a 2D convolution during export.

This does not magically turn an MLP into a convolutional network at runtime. It annotates a particular export-layer index with a conv interpretation so that:

• The exported graph uses conv-shaped tensors/operators, and
• Import can re-attach pooling/flatten metadata appropriately.

Pitfall: mappings must match the actual layer sizes. If inHeight * inWidth * inChannels does not correspond to the prior layer width (and similarly for outputs), export or import may reject the model.

OnnxAttribute

ONNX node attribute payload.

This simplified JSON-first shape is enough for the operators emitted by the current exporter. It intentionally avoids protobuf-level complexity while still preserving the attribute variants needed by the importer.

OnnxDimension

One dimension inside an ONNX tensor shape.

Use dim_value for fixed numeric widths and dim_param for symbolic names such as a batch dimension.

OnnxGraph

Graph body of an ONNX-like model.

The exporter writes three main collections here:

• inputs and outputs describe graph boundaries,
• initializer stores constant tensors such as weights and biases,
• node stores the ordered operator payloads that consume those tensors.

OnnxMetadataProperty

Canonical metadata key-value pair used in ONNX model metadata_props.

OnnxModel

ONNX-like model container (JSON-serializable).

This is the main “wire format” object in this folder. Persist it as JSON text:

const jsonText = JSON.stringify(model);
const restoredModel = JSON.parse(jsonText) as OnnxModel;

Notes:

• metadata_props contains NeatapticTS-specific keys (layer sizes, recurrent flags, conv/pool mappings, etc.). This is where most round-trip hints live.
• Initializers currently store floating-point weights in float_data.

Security/trust boundary:

• Treat this as untrusted input if it comes from outside your process.

OnnxNode

One ONNX operator invocation inside the graph.

Nodes connect named tensors rather than object references, which keeps the exported payload easy to serialize, inspect, and diff as plain JSON.

OnnxShape

ONNX tensor type shape.

OnnxTensor

Serialized tensor payload stored inside graph initializers.

NeatapticTS currently writes floating-point parameter vectors and matrices to float_data, along with the tensor name, element type, and logical shape.

OnnxTensorType

ONNX tensor type.

OnnxValueInfo

ONNX value info (input/output description).

Pool2DMapping

Mapping describing a pooling operation inserted after a given export-layer index.

This is represented as metadata and optional graph nodes during export. Import uses it to attach pooling-related runtime metadata back onto the reconstructed network (when supported).