tf2tfjs — TensorFlow to TensorFlow.js Conversion Service

A web service that converts TensorFlow SavedModel files to TensorFlow.js format. Built for Windows users who cannot directly use tensorflowjs_converter, and designed as an asynchronous microservices architecture using RabbitMQ.

Upload a .zip containing your TensorFlow SavedModel, and the service will convert it and provide a downloadable TensorFlow.js graph model.

Architecture

User ──▶ Frontend (Nginx) ──▶ API (Flask) ──▶ RabbitMQ ──▶ Consumer ──▶ convert.sh
                                   │                                        │
                                   └──────── MongoDB ◀──────────────────────┘

Service	Description
Frontend	Nginx-served static page for uploading models and downloading results
API	Flask REST API handling uploads, status queries, and file downloads
Consumer	Background worker that picks jobs from RabbitMQ and runs the conversion
RabbitMQ	Message broker decoupling uploads from processing
MongoDB	Stores conversion status and model metadata

API Endpoints

Method	Endpoint	Description
POST	/upload	Upload a .zip file containing a TensorFlow SavedModel. Returns a model_id.
GET	/status?model_id=<id>	Check conversion status. Returns status: 0 (success) or status: 1 (failure).
GET	/get_model?model_id=<id>	Download the converted model as a .tar.gz archive.

Getting Started

Prerequisites

• Docker
• Docker Compose

Build

docker build -t tfjs-api:latest -f api/Dockerfile api/
docker build -t tfjs-frontend:latest -f page/Dockerfile page/

Run (Production)

docker-compose up

Service	URL
Frontend	http://localhost:5001
API	http://localhost:5000
Mongo Express	http://localhost:8080

Run (Development)

docker-compose -f docker-compose-dev.yaml up

Development mode additionally exposes RabbitMQ management (port 15672) and a Jupyter Lab instance (port 4999).

Usage

1. Open the frontend at http://localhost:5001.
2. Select a .zip file containing a TensorFlow SavedModel (must include .pb files and a variables/ folder).
3. Click Upload and Convert.
4. Wait for the conversion to complete — the page polls automatically.
5. Click Download to get the converted TensorFlow.js model as a .tar.gz archive.

Configuration

All service configuration lives in api/configuration.yaml:

Setting	Default	Description
RabbitMQ host	rabbitmq	Docker service hostname
MongoDB host	mongodb	Docker service hostname
MongoDB database	tf2tfjs	Database name
API port	5000	Flask/Gunicorn listen port

Tech Stack

• Backend: Python 3.10, Flask, Gunicorn, Pika, PyMongo
• Conversion: tensorflowjs_converter (tf_saved_model → tfjs_graph_model)
• Frontend: HTML, CSS, JavaScript, Nginx
• Infrastructure: Docker, Docker Compose, RabbitMQ, MongoDB