SavedModel

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

const graph = parseSavedModel(bytes); // synchronous, returns ModelGraph

Or use the unified entry point - .pb files are routed here automatically:

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

const graph = await parseModel(bytes, "model.pb");

Format variants#

keras_metadata.pb#

Detected by first byte 0x0a (protobuf field 1, length-delimited). Contains a Keras layer graph serialized as JSON inside the protobuf.

• Supported topologies: Sequential, Functional
• InputLayer entries -> ModelGraph.inputs (excluded from nodes)
• Layer class_name -> node opType
• Layer config fields -> node attributes
• Edges resolved from inbound_nodes for Functional; chained for Sequential

saved_model.pb#

Detected by first byte 0x08 (protobuf field 1, varint - schema version). Contains a TensorFlow GraphDef with raw TF ops.

• Placeholder nodes -> ModelGraph.inputs
• All other ops -> ModelGraph.nodes
• Output nodes inferred as nodes whose outputs are never consumed as inputs
• Const node tensor shape and dtype are folded into ModelGraph.initializers; consumers see them as initializer edges in the rendered graph
• StatefulPartitionedCall and PartitionedCall function bodies are inlined from the library.function table; body node names are prefixed with the call-site name to avoid collisions
• VarHandleOp nodes mark variables backed by an external checkpoint - the parser sets ModelGraph.hasExternalWeights = true so a host app knows to load them via loadSavedModelWeights

Graph structure (saved_model.pb)#

Placeholder (input)
  ↓
Conv2D ← Const (weight)
  ↓
BiasAdd ← Const (bias)
  ↓
Relu
  ↓
...
  ↓
Softmax (output)

Notes#

• parseModel detects .pb by filename extension; the first-byte check then selects the variant.
• ModelGraph.weights is not populated by parseSavedModel itself. For TF2 models, load the checkpoint pair (variables.index + variables.data-00000-of-00001) with loadSavedModelWeights and call attachCheckpointToGraph to produce a graph with weights re-keyed by node name. See Weights.
• Control dependencies (inputs prefixed with ^) are ignored.
• Port suffixes (:0, :1) are stripped from input tensor names. As a consequence, multi-output ops (Split, TopK, …) cannot be fully disambiguated - every consumer of split:0, split:1 is recorded as a consumer of split. Connectivity is preserved; per-port labelling is not.
• Each node is recorded with a single output named after the node itself. Multi-output TF ops are visualised as if they had one fan-out.
• Non-fatal per-node errors are attached as warnings on the returned graph.
• Throws ParseError if the file is too short or has unrecognized first byte.