StarDistをDockerで動かす

Stardist GPU環境をDockerで構築する（Ubuntu + TensorFlow + Jupyter + StarDist）

環境：

• Ubuntu 20.04 LTS
• NVIDIA RTX 4000 SFF Ada
• NVIDIA Container Toolkitインストール済
• Docker, Docker Composeインストール済

---

Dockerfile

# Use TensorFlow 2.10.1 GPU Jupyter image as base
FROM tensorflow/tensorflow:2.10.1-gpu-jupyter

# Ensure root privileges
USER root

# System updates and tools
RUN apt-get update && \
    apt-get install -y sudo gosu && \
    rm -rf /var/lib/apt/lists/*

# Upgrade pip
RUN pip3 install --upgrade pip

# Install Python packages
RUN pip3 install \
    "numpy<2" \
    "stardist[tf2]" \
    gputools==0.2.15 \
    EDT \
    reikna==0.8.0 \
    numba \
    jupyterlab

# Setup working directory and symlinks
RUN mkdir -p /notebooks && \
    ln -s /data /notebooks/data && \
    ln -s /tf/tensorflow-tutorials /notebooks/tensorflow-tutorials

# Copy helper script
COPY createuser.sh /usr/bin/createuser.sh
RUN chmod +x /usr/bin/createuser.sh

# Expose port for Jupyter Lab
EXPOSE 8888

# Use custom user creation script as entrypoint
ENTRYPOINT ["/usr/bin/createuser.sh"]

# Default command
CMD ["jupyter", "lab", "--ip=0.0.0.0", "--notebook-dir=/notebooks", "--no-browser", "--allow-root"]

---

docker-compose.yml

version: '3.8'
services:
  stardist:
    build:
      context: .
      dockerfile: Dockerfile
    container_name: stardist_container
    runtime: nvidia
    environment:
      - USER_NAME=dockeruser
      - GROUP_NAME=dockeruser
      - USER_ID=1000
      - GROUP_ID=1000
      - USER_PASSWORD=yourpassword
      - ROOT_PASSWORD=yourrootpassword
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
    volumes:
      - /home/youruser/Data:/data
    ports:
      - "8777:8888" #自分の環境では8888がもう使われているので。
    tty: true

---

createuser.sh

#!/bin/bash
USER_NAME=${USER_NAME:-dockeruser}
GROUP_NAME=${GROUP_NAME:-dockeruser}
USER_ID=${USER_ID:-1000}
GROUP_ID=${GROUP_ID:-1000}
USER_PASSWORD=${USER_PASSWORD:-password}
ROOT_PASSWORD=${ROOT_PASSWORD:-rootpassword}

# Group setup
if ! getent group "$GROUP_NAME" >/dev/null 2>&1; then
  groupadd -g "$GROUP_ID" "$GROUP_NAME"
fi

# User setup
if ! id -u "$USER_NAME" >/dev/null 2>&1; then
  useradd -m -s /bin/bash -u "$USER_ID" -g "$GROUP_NAME" "$USER_NAME"
  echo "$USER_NAME:$USER_PASSWORD" | chpasswd
  usermod -aG sudo "$USER_NAME"
fi

echo "root:$ROOT_PASSWORD" | chpasswd

# Drop privileges to created user
exec gosu "$USER_NAME" "$@"

---

entrypoint.sh

#!/bin/bash
# Print the URL to access Jupyter Lab
echo "Access Jupyter notebook at http://localhost:8888 (or http://<your_docker_host>:8888 if accessing from another machine)"

# Start Jupyter Lab
exec jupyter lab --ip=0.0.0.0 --no-browser --allow-root

---

これを Stardist ディレクトリにおいてdocker compose up --build -d

💡 createuser.shとentrypoint.shは、「rootでログインするのがなんとなく嫌」「ボリュームマウントをミスなく行いたい」という目的でおまじない的に入れています。
💡notebooksディレクトリを作っておかないとJupyter Labを立ち上げたときにマウントしたボリュームが見えないです。

参考：
https://forum.image.sc/t/difficulty-installing-stardist-tensorflow-gputools-with-anaconda/104305/15