Custom YOLO Everywhere

Custom YOLO Everywhere is an open-source solution that bridges the gap between image annotation and custom object detection model deployment. This project seamlessly integrates Label Studio with a YOLO client, automating the entire workflow from annotation export to model training and multi-format export. Whether you want to deploy on the web using TensorFlow.js, on mobile, or other native platforms via TFLite, this tool empowers developers, researchers, and hobbyists to build and deploy custom YOLO models.

Video Tutorial 🎥

For a step-by-step guide on setting up and using Custom YOLO Everywhere, check out our tutorial:
📺 Watch the video on YouTube

Prerequisites

• Docker & Docker Compose: to run Label Studio.
• Python (3.10+): for training and exporting the YOLO model

Docker Setup

This project uses Docker Compose to manage the services:

Label Studio: Runs on port 8080.

.env File

Create a .env file in the root with your environment variables:

LABEL_STUDIO_URL=http://localhost:8080

Note: The env.template can be used as reference

Running the Containers

1. Build and start the containers:

docker-compose up --build -d

Label Studio will be available at http://localhost:8080. Use it to annotate your images. When you export your annotations to YOLO format from Label Studio, save the files into the yolo-custom/dataset/train and yolo-custom/dataset/val directories (or run an automation script, see below).

3. Running the Interactive Client

Install the dependencies

pip install -r requirements.txt

Now you can run the interactive client from the project root by executing:

python app/client.py

Command-Line overview

This project includes an interactive command-line client that simplifies managing configurations and running key tasks. Once you launch the client, you will see a menu with several options:

Create a New Configuration:

Launch an interactive process to create a new configuration file. This process will prompt you for your Label Studio API key (you can find it under the "account & settings" sections of your Label Studio installation), then you'll be able to select a project and its available tags, and ask for additional training parameters (e.g., train ratio and epochs). The generated configuration is saved in the configs/ directory.

Prepare Dataset:

Lists all configuration files found in the configs/ directory. After you select one, the client calls the dataset preparation script (app/prepare.py) using that configuration. This process exports annotations from Label Studio, merges multiple exports if necessary, splits the dataset into training and validation sets, and generates a proper data.yaml file.

Train Based on Configuration:

Lists all configuration files from the configs/ directory and allows you to select one to use for training. Once selected, it triggers the training script (app/train.py) with the chosen configuration as a parameter.

Export Model to TensorfFlow Formats

Similarly, this option lists the configuration files. When you select one, it runs the export script (app/export.py) with the chosen configuration file as a parameter to convert your trained model to TensorFlow format.

Running on Apple Silicon

To run on Apple Silicon (M1/M2), create and activate an ARM-native virtual environment and install the required packages:

conda create -n arm-env python=3.10
conda activate arm-env

pip install --upgrade pip setuptools wheel
pip install -r requirements.txt
python app/client.py

M1 troubleshooting

Sometimes, the JAX library presents issues on Apple devices. Follow the official tutorial to ensure a proper JAX installation.

Then, install an ARM-compatible version of JAX (CPU-only) from Google’s storage:

python -m pip install jax-metal

Test it:

python -c 'import jax; print(jax.numpy.arange(10))'