Docker

Prebuilt images with docTR installed are published on the GitHub Container Registry. They are based on CUDA 12.2 and ready for GPU inference.

Run an image

docker run -it --gpus all ghcr.io/mindee/doctr:<tag> bash

Drop --gpus all to run on the CPU. For GPU support, your host needs:

• an NVIDIA driver supporting CUDA 12.2 or newer, otherwise PyTorch cannot initialize the GPU;

• the NVIDIA Container Toolkit so that Docker can access the GPU.

Tags

Tags follow the pattern <deps>-py<python_version>-<version>:

<deps>

torch for the core package, or torch-viz-html-contrib with all the optional extras.

<python_version>

The Python version of the image, e.g. 3.11.15, 3.12.9 or 3.13.14.

<version>

A release tag (e.g. v1.1.0), or a YYYY-MM date for the images rebuilt from the main branch every three months.

For instance, torch-viz-html-contrib-py3.11.15-v1.1.0 is release 1.1.0 with all the extras on Python 3.11.15. Browse the registry for the available tags.

Build your own image

The Dockerfile at the root of the repository accepts build arguments to choose the Python version, the extras and the docTR version:

docker build -t doctr \
  --build-arg FRAMEWORK="torch,viz,html,contrib" \
  --build-arg PYTHON_VERSION=3.11.15 \
  --build-arg DOCTR_VERSION=v1.1.0 \
  .

Argument	Default	Description
FRAMEWORK	torch	Extras to install, comma-separated (e.g. torch,viz,html)
PYTHON_VERSION	3.11.15	Python version, compiled from source
DOCTR_VERSION	main	Git branch, tag or commit of docTR to install

Tip

Pretrained weights are downloaded at runtime. To avoid it, warm the cache when building the image, e.g. with RUN python -c "from doctr.models import ocr_predictor; ocr_predictor(pretrained=True)", or mount a volume on the cache directory (see Configuration).