@wetron/savedmodel

TF SavedModel parser for wetron. Reads .pb files in two formats:

• saved_model.pb - TensorFlow SavedModel GraphDef (raw TF ops)
• keras_metadata.pb - Keras layer metadata protobuf

Returns a ModelGraph IR. Graph structure only - no weight tensors are deserialized.

Install

pnpm add @wetron/savedmodel

Included automatically when you install @wetron/core.

API

import { parseSavedModel } from "@wetron/savedmodel";

const bytes = new Uint8Array(await file.arrayBuffer());
const graph = parseSavedModel(bytes); // synchronous, returns ModelGraph

Throws ParseError from @wetron/common/ir if the file is too short or has unrecognized .pb content.

Checkpoint loading

Variable weights live outside the .pb file in the TF2 variables/ checkpoint pair. Load and attach them with:

import {
  loadSavedModelWeights,
  loadSavedModelWeightsFromUrls,
  attachCheckpointToGraph,
} from "@wetron/savedmodel";

// From local files
const loaded = await loadSavedModelWeights(indexFile, dataFile);

// From URLs (one URL per shard, in shard order)
const loaded = await loadSavedModelWeightsFromUrls(
  "https://.../variables.index",
  "https://.../variables.data-00000-of-00001",
);

const graphWithWeights = attachCheckpointToGraph(graph, loaded);

graph.hasExternalWeights is true whenever at least one VarHandleOp is present, signalling the host app to prompt for the checkpoint files.

Format detection

The first byte determines which variant is parsed:

| First byte | Format | | ---------- | ---------------------------------- | | 0x0a | keras_metadata.pb (Keras layers) | | 0x08 | saved_model.pb (TF GraphDef) |

parseModel from @wetron/core routes .pb files here automatically based on filename.

What gets parsed

keras_metadata.pb

• Layer graph from the embedded keras_metadata JSON
• Supported topologies: Sequential, Functional
• InputLayer entries -> ModelGraph.inputs
• Layer class_name -> node opType
• Layer config fields -> node attributes

saved_model.pb

• First MetaGraphDef -> GraphDef nodes
• Placeholder nodes -> ModelGraph.inputs
• All other ops -> ModelGraph.nodes
• Output nodes inferred as nodes whose outputs are never consumed as inputs

Notes

• ModelGraph.initializers is always empty - weight data is not parsed.
• Const nodes (weight constants) appear as graph nodes with category constant.
• Control dependencies (inputs prefixed with ^) are ignored.
• Port suffixes (:0, :1) are stripped from input tensor names.
• Non-fatal per-node errors are attached as warnings on the returned graph.