Install

These guides cover the stable release (v1.1.2). If you are installing v2.0.0-beta.4, follow the Beta Guides instead. The beta does not use Google Chrome, and the procedure differs.

YORU can be installed in two ways:

• Install with uv builds the whole environment, including Python itself, with one command on Windows, Linux and macOS. This is the only supported route on macOS.
• Install with conda is the original route, for Windows and Linux with an NVIDIA GPU.

Whichever you choose, read Prerequisites first. Neither uv nor conda can install these for you.

Prerequisites

1. A Chromium browser (all platforms)

The YORU launcher is a web page served by Eel, which opens it in Google Chrome or Chromium. Eel looks for the browser here:

If no browser is found, YORU stops at startup with:

EnvironmentError: Can't find Google Chrome/Chromium installation

Microsoft Edge is built on Chromium, but Eel does not find it. Safari is not supported. Only the launcher needs the browser: the training, evaluation, analysis and real-time windows are native windows.

2. An NVIDIA driver, to use a GPU (Windows / Linux)

Install the GPU driver. It has to support CUDA 12.x: 527.41 or newer on Windows, 525.60.13 or newer on Linux. Check the installed version with:

nvidia-smi

The CUDA toolkit is not required by either route. The PyTorch wheels carry their own CUDA runtime (cudart, cuBLAS and cuDNN are inside the torch package). Install the toolkit only if you need nvcc to compile CUDA extensions of your own, or want the profiling tools.

The CUDA Version that nvidia-smi reports is the highest version your driver supports, not the version in use. A driver reporting 13.x runs the CUDA 12.4 wheels without a toolkit installed.

Without a usable GPU, YORU runs on the CPU. Everything works, but more slowly.

3. macOS extras (Apple Silicon)

• YORU needs macOS 14 (Sonoma) or later on Apple Silicon. Intel Macs are not supported, because the macOS wheels of the PyTorch and Qt versions YORU pins are arm64 only.

• Install the Xcode Command Line Tools before uv sync. Two dependencies (imgui and gevent) have no arm64 wheels and are compiled during the sync:

    xcode-select --install
  

• macOS asks for permission the first time YORU uses a camera (Camera), the key-press triggers (Input Monitoring / Accessibility, via pynput) or screen capture (Screen Recording, via mss). The prompts are for the terminal application you launched YORU from. Grant them in System Settings > Privacy & Security, then relaunch.
• The NI-DAQ closed-loop path needs a driver that NI ships for Windows only, so it is not available on macOS.

Install with uv

uv builds the environment from the pyproject.toml and uv.lock in the repository. It picks the right PyTorch build for your platform: the CUDA wheels on Windows and Linux, and the MPS-enabled build on macOS. You do not need a system Python or conda, because uv downloads Python 3.10 for the project.

1. Install uv.

   Windows (PowerShell):

    winget install --id=astral-sh.uv -e
   

   macOS / Linux:

    curl -LsSf https://astral.sh/uv/install.sh | sh
   

   brew install uv also works on macOS. The uv installation guide lists the other installers. Open a new terminal afterwards so that uv is on PATH.

2. Download or clone the YORU project.

    cd "Path/to/download"
    git clone https://github.com/Kamikouchi-lab/YORU.git
   

   If you do not have git, you can download the ZIP from GitHub instead.

3. Build the environment.

    cd YORU
    uv sync
   

   The first sync downloads PyTorch and takes a few minutes.

4. Run YORU from the repository root. The launcher looks for web/ and config/ relative to the working directory.

    uv run yoru
   

   uv run python -m yoru does the same thing. There is no uv init step: the project is already initialised.

To check that the GPU was picked up:

uv run python -c "import torch; print(torch.cuda.is_available())"

Install with conda

1. Check that Google Chrome is installed (see Prerequisites).

2. Check the installation of Miniconda.

   Anaconda’s TERMS OF SERVICE was changed. If you use Anaconda in an organization that has two hundred (200) or more employees or contractors, you have to be careful.

   Currently, you can use miniconda freely.

3. Download or clone the YORU project.

   a. Install git

    conda install git
   

   b. Clone the repository

    cd "Path/to/download"
    git clone https://github.com/Kamikouchi-lab/YORU.git
   

4. Install the GPU driver. The CUDA toolkit is not needed: the PyTorch wheels in step 7 carry their own CUDA runtime, so the cu118 / cu121 choice there only has to be one that your driver supports.

5. Create a virtual environment using YORU.yml in the command prompt or Anaconda prompt.

     conda env create -f "Path/to/YORU.yml"
   

   The environment file pins Python 3.10 (v1.1.1 used Python 3.9). If you are upgrading from v1.1.1, recreate the environment: conda env remove -n yoru, then create it again.

6. Activate the virtual environment in the command prompt or miniconda prompt.

     conda activate yoru
   

7. Install PyTorch for a CUDA version that your driver supports.

   • For CUDA==11.8

    pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
   

   • For CUDA==12.1

    pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121
   

   (torch, torchvision and torchaudio will be installed.)

   RTX 50-series (Blackwell) cards need a newer build. See working-example.md in the repository, which records a working RTX 5070 Ti setup on torch 2.8.0+cu128.

8. Run YORU in a command prompt or miniconda prompt.

    conda activate yoru
    cd "Path/to/YORU/project/folder"
    python -m yoru
   

Choosing the compute device

YORU picks its compute device automatically, in this order: CUDA, then Apple MPS, then the CPU.

To choose the device yourself, set the YORU_DEVICE environment variable before launching (cuda, mps, cpu, or a CUDA index such as 0), or use the Device selector in the Training GUI:

YORU_DEVICE=cpu uv run yoru

On Windows, run set YORU_DEVICE=cpu before the launch command.

• YORU_DEVICE applies to YOLOv8, YOLO11, RT-DETR and the torchvision detectors. YOLOv5 inference loads through torch.hub, which always uses CUDA when it is available and the CPU otherwise.
• If the requested device is not available, YORU falls back to the next best one and writes a warning to the log file (see below).
• On Apple Silicon, MPS clearly speeds up training, but it is not faster than the CPU for real-time inference with the small YOLO models (60.8 FPS on MPS against 68.7 FPS on CPU for yolov8n at 640×480). For real-time detection on a Mac, try YORU_DEVICE=cpu.
• Faster R-CNN / Mask R-CNN / SSD need torchvision 0.29 or newer to train on MPS. The uv environment pins a new enough build. With an older one, YORU falls back to the CPU and shows a message.

Log files

YORU writes runtime errors and work logs to a user directory:

The log holds the full traceback of every error that the GUIs catch, so attach it when you report a problem on GitHub Issues. Set the YORU_HOME environment variable to move the directory somewhere else.

Next