Wetron
Wetron

parseModel / detectFormat

parseModel

import { parseModel } from "@wetron/core";

async function parseModel(bytes: Uint8Array, filename?: string): Promise<ModelGraph>;

Inspects the buffer’s magic bytes (with filename extension as a tiebreaker) to identify the format, then dynamically imports and calls the matching parser package.

Throws ParseError if the format cannot be detected or parsing fails.

Direct parser imports

Call parsers directly to avoid the auto-detection overhead, or if you want to exclude specific parsers from your bundle:

import { parseOnnx } from "@wetron/onnx";
import { parseTflite } from "@wetron/tflite";
import { parseKeras } from "@wetron/keras";
import { parseTorchscript } from "@wetron/torchscript";
import { parseExecutorch } from "@wetron/executorch";
import { parseSavedModel } from "@wetron/savedmodel";

const graph = parseOnnx(bytes);
const graph = parseTflite(bytes);
const graph = parseKeras(bytes);
const graph = parseTorchscript(bytes);
const graph = parseExecutorch(bytes);
const graph = parseSavedModel(bytes);

detectFormat

import { detectFormat } from "@wetron/core";

type Format = "onnx" | "tflite" | "keras" | "torchscript" | "executorch" | "savedmodel" | "unknown";

function detectFormat(bytes: Uint8Array, filename?: string): Format;

Returns a format string - never throws. Useful for showing format badges in a UI before parsing.

FormatDetection
savedmodel.pb filename extension (checked before ONNX)
onnxprotobuf field 1 varint tag 0x08
tfliteTFL3 or ODLF at offset 4
kerasZIP magic PK\x03\x04 + .keras extension (default for ZIP)
torchscriptZIP magic PK\x03\x04 + .pt/.ptl extension, or PTMF at 4
executorchET12 at offset 4
unknownno match

modelGraphToFlow

import { modelGraphToFlow } from "@wetron/core";

function modelGraphToFlow(graph: ModelGraph): { nodes: FlowNode[]; edges: FlowEdge[] };

Converts a ModelGraph to layout-positioned FlowNode[] and FlowEdge[] ready for ReactFlow or SvelteFlow. Dagre is applied top-to-bottom. Call directly only when building a custom renderer; the renderer packages call it automatically.

opCategory

import { opCategory } from "@wetron/core";

type OpCategory =
  | "input"
  | "output"
  | "conv"
  | "activation"
  | "normalization"
  | "pooling"
  | "reshape"
  | "math"
  | "reduction"
  | "merge"
  | "attention"
  | "recurrent"
  | "quantization"
  | "constant"
  | "logic"
  | "unknown";

function opCategory(opType: string): OpCategory;

Maps an op type string to a semantic category. Used by the renderer to assign node colours.

opInputLabels

import { opInputLabels } from "@wetron/core";

function opInputLabels(opType: string): readonly string[];

Returns named input slot labels for known ops (e.g. Conv -> ["X", "W", "B"]). Returns an empty array for unrecognised ops.

filterGraph

import { filterGraph } from "@wetron/core";

function filterGraph(graph: ModelGraph, query: string): ReadonlySet<string>;

Returns the set of node names whose opType or name contains query (case-insensitive). Empty query returns an empty set. Used by the renderer’s search box to dim non-matching nodes.

Prev
ONNX
Next
React