Keras

import { parseKeras } from "@wetron/keras";

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

What is parsed#

• config.json from inside the .keras ZIP archive - contains the full layer graph
• Supported class_name values: Sequential, Functional
• InputLayer entries -> ModelGraph.inputs (excluded from nodes)
• Layer class_name -> node opType
• Layer config fields -> node attributes (name, dtype, trainable filtered out)
• Edges resolved from inbound_nodes[].args[].keras_history for Functional models; chained sequentially for Sequential models

Keras 3 tensor references#

Keras 3 serializes tensor references as { class_name: "__keras_tensor__", config: { keras_history: [...] } }. The parser handles both this form and the older { keras_history: [...] } shorthand.

Notes#

• Uses fflate for ZIP decompression.
• ModelGraph.initializers is empty and ModelGraph.weights is absent - Keras 3 stores trained weights in a separate .weights.h5 file which this parser does not read.
• ModelGraph.tensorShapes is populated from InputLayer batch shapes only (null batch dimension -> -1).
• Multi-input merge layers (e.g. Concatenate) correctly receive multiple inputs entries from the args array in inbound_nodes.
• Throws ParseError if config.json is missing, invalid JSON, or uses an unsupported model class.