Implementation:Tensorflow Tfjs Tensorflowjs Converter CLI

Knowledge Sources	TensorFlow.js tfjs-converter/README.md TensorFlow.js Conversion Guide
Domains	Model_Persistence, Deployment
Principle	Principle:Tensorflow_Tfjs_Model_Format_Conversion
Type	External Tool Doc
Last Updated	2026-02-10 00:00 GMT

Environment:Tensorflow_Tfjs_Python_Converter

Overview

This implementation documents the tensorflowjs_converter command-line interface, the primary tool for converting Python TensorFlow models into TensorFlow.js format. The CLI reads models from various source formats and produces the model.json manifest plus .bin weight shard files that TensorFlow.js can load in the browser or Node.js.

Source Reference

tfjs-converter/README.md:L87-171 — The CLI usage, parameter definitions, and conversion examples.

CLI Syntax

tensorflowjs_converter \
    --input_format=<INPUT_FORMAT> \
    --output_format=<OUTPUT_FORMAT> \
    [OPTIONS] \
    <INPUT_PATH> \
    <OUTPUT_PATH>

Parameters

Required Parameters

Parameter	Description	Values
<INPUT_PATH>	Path to the source model (directory, file, or URL)	File path or URL string
<OUTPUT_PATH>	Directory where the converted model will be written	Directory path (created if it does not exist)
--input_format	The format of the source model	tf_saved_model, keras, keras_saved_model, tf_hub, tfjs_layers_model, tf_frozen_model

Output Format

Parameter	Description	Values	Default
--output_format	The target TF.js format	tfjs_graph_model, tfjs_layers_model	Depends on input format

SavedModel-Specific Parameters

Parameter	Description	Default
--signature_name	The serving signature to export from the SavedModel	serving_default
--saved_model_tags	Comma-separated MetaGraphDef tags	serve

Frozen Graph Parameters

Parameter	Description	Default
--output_node_names	Comma-separated list of output node names	Required for frozen graphs

Quantization Parameters

Parameter	Weight Precision	Size Reduction	Accuracy Impact
--quantize_float16	float32 to float16	~50%	Minimal
--quantize_uint8	float32 to uint8	~75%	Low to moderate
--quantize_uint16	float32 to uint16	~50%	Low

Other Parameters

Parameter	Description	Default
--weight_shard_size_bytes	Maximum byte size of each weight shard file	4194304 (4 MB)
--control_flow_v2	Enable TF Control Flow V2 ops	True
--strip_debug_ops	Remove Assert and debug operations from the graph	True
--skip_op_check	Skip the unsupported op check (use with caution)	False
--metadata	Custom metadata JSON string to embed in model.json	None

Inputs and Outputs

Inputs

TensorFlow SavedModel directory (containing saved_model.pb + variables/)
Keras HDF5 file (.h5 extension)
Keras SavedModel directory (Keras model saved in SavedModel format)
TF Hub module URL (https://tfhub.dev/...)
TF Frozen Graph (.pb file with frozen weights)
TF.js Layers Model directory (for reverse conversion or format change)

Outputs

The converter writes the following files to <OUTPUT_PATH>:

model.json: The model manifest containing:
- modelTopology: The computation graph (graph model) or Keras layer config (layers model)
- weightsManifest: Array of weight group descriptors with shard filenames and tensor metadata
- format: Model format identifier
- generatedBy: Converter version string
- convertedBy: TF.js version string
group1-shard1ofN.bin, group1-shard2ofN.bin, ...: Binary weight shard files

Conversion Examples

SavedModel to Graph Model

# Convert a TensorFlow SavedModel to TF.js graph model
tensorflowjs_converter \
    --input_format=tf_saved_model \
    --output_format=tfjs_graph_model \
    --signature_name=serving_default \
    --saved_model_tags=serve \
    /path/to/saved_model \
    /path/to/tfjs_output

SavedModel to Graph Model with Quantization

# Convert with uint8 quantization for ~75% size reduction
tensorflowjs_converter \
    --input_format=tf_saved_model \
    --output_format=tfjs_graph_model \
    --quantize_uint8 \
    /path/to/saved_model \
    /path/to/tfjs_output_quantized

Keras HDF5 to Layers Model

# Convert a Keras HDF5 model to TF.js layers model
tensorflowjs_converter \
    --input_format=keras \
    --output_format=tfjs_layers_model \
    /path/to/model.h5 \
    /path/to/tfjs_output

Keras SavedModel to Layers Model

# Convert a Keras model saved in SavedModel format
tensorflowjs_converter \
    --input_format=keras_saved_model \
    --output_format=tfjs_layers_model \
    /path/to/keras_saved_model \
    /path/to/tfjs_output

TF Hub Module to Graph Model

# Convert a TF Hub module directly from URL
tensorflowjs_converter \
    --input_format=tf_hub \
    --output_format=tfjs_graph_model \
    'https://tfhub.dev/google/imagenet/mobilenet_v2_100_224/classification/5' \
    /path/to/tfjs_output

Frozen Graph to Graph Model

# Convert a frozen graph with explicit output nodes
tensorflowjs_converter \
    --input_format=tf_frozen_model \
    --output_format=tfjs_graph_model \
    --output_node_names='softmax_output' \
    /path/to/frozen_model.pb \
    /path/to/tfjs_output

Float16 Quantization with Custom Shard Size

# Convert with float16 quantization and 1MB shards
tensorflowjs_converter \
    --input_format=tf_saved_model \
    --output_format=tfjs_graph_model \
    --quantize_float16 \
    --weight_shard_size_bytes=1048576 \
    /path/to/saved_model \
    /path/to/tfjs_output

Reverse Conversion: TF.js to Keras

# Convert a TF.js layers model back to Keras HDF5
tensorflowjs_converter \
    --input_format=tfjs_layers_model \
    --output_format=keras \
    /path/to/tfjs_model/model.json \
    /path/to/keras_output.h5

Output Verification

After conversion, verify the output files:

# List the converted output files
ls -la /path/to/tfjs_output/
# Expected: model.json, group1-shard1ofN.bin, ...

# Check model.json structure
python -c "
import json
with open('/path/to/tfjs_output/model.json') as f:
    model = json.load(f)
print('Format:', model.get('format'))
print('Generated by:', model.get('generatedBy'))
print('Weight groups:', len(model.get('weightsManifest', [])))
for group in model.get('weightsManifest', []):
    print('  Shards:', group.get('paths'))
    print('  Weights:', len(group.get('weights', [])))
"

Conversion Format Matrix

Source Format	--input_format	--output_format	Typical Use Case
TF SavedModel	tf_saved_model	tfjs_graph_model	Production inference-only deployment
TF SavedModel	tf_saved_model	tfjs_layers_model	When layers API is needed in JS
Keras SavedModel	keras_saved_model	tfjs_layers_model	Keras models with fine-tuning support
Keras HDF5	keras	tfjs_layers_model	Legacy Keras model files
TF Hub	tf_hub	tfjs_graph_model	Pre-trained models from TF Hub
Frozen Graph	tf_frozen_model	tfjs_graph_model	Legacy TF 1.x frozen models
TF.js Layers	tfjs_layers_model	keras	Export back to Python Keras
TF.js Layers	tfjs_layers_model	keras_saved_model	Export back to Python SavedModel

Common Issues

Issue	Cause	Resolution
Unsupported op: OpName	The model uses a TensorFlow operation not implemented in TF.js	Check the supported ops list; consider rewriting the model to avoid unsupported ops, or use --skip_op_check (inference will fail at runtime for those ops)
No signature found	SavedModel has no serving signature	Export the model with explicit signatures; see Implementation:Tensorflow_Tfjs_Tf_Saved_Model_Save
Memory error during conversion	Very large model exceeds system memory	Increase system swap space; use a machine with more RAM
Output node not found	Incorrect --output_node_names for frozen graph	Inspect the graph with saved_model_cli or TensorBoard to find correct node names
Weight shard too large	Default 4MB shard size may be too large or small	Adjust --weight_shard_size_bytes for your deployment needs

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment