Skip to content

INSTALLATION.md

Latest commit

 

History

History
134 lines (82 loc) · 3.9 KB

INSTALLATION.md

File metadata and controls

134 lines (82 loc) · 3.9 KB

Installation

Choose one of two primary paths:

  • Docker - Recommended for maximum compatibility across different Python environments
  • Local server (no Docker) - Run the server directly with uv

Advanced/optional modes (for experienced users) are listed later in this file.

Docker (Recommended)

Requires Docker with Compose.

Build and start locally

# Use pre-built Docker Hub image (default)
docker compose up -d

# Build locally from this repository
docker compose -f docker-compose.local.yml up -d --build

(Your system might use docker-compose instead)

docker-compose.yml is now the Docker Hub default. docker-compose.local.yml is an override for local builds.

For reproducible deployments, pin a version tag in docker-compose.yml (for example docker.io/docminus/decimer_api:1.2.0) instead of :latest.

Install Python client

To access via the Python API, install the client package into your existind/desired environment:

pip install ./packages/decimerapi/

This is not required if you want to use plain HTTP requests.

GPU Notes

  • Docker doesn't support Mac GPU (Docker is Linux-only)
  • If GPU runtime is unavailable, TensorFlow falls back to CPU automatically
  • Linux/Windows GPU support depends on host NVIDIA/CUDA runtime setup
  • For build with Nvidia GPU, a CUDA based docker build option is available.
    • nvidia-smi on the host should show GPU status without errors

To build and run with Nvidia GPU support:

docker compose -f docker-compose.cuda.yml up -d --build

Note: this hasn't been tested on all possible platforms/configurations.

Local Installation (No Docker)

Requirements

  • Python >= 3.10 (3.10 recommended on Mac; newer versions not guaranteed to work)
  • uv package manager

Install uv:

  • macOS: brew install uv or pip install uv
  • Linux: curl -LsSf https://astral.sh/uv/install.sh | sh
  • Windows: irm https://astral.sh/uv/install.ps1 | iex

Or simply: pip install uv in the existing environment

Quick Install (All Platforms)

uv sync
uv run python decimer_server.py

This creates a .venv using Python 3.1x and installs all dependencies (including decimer_image_classifier and decimerapi packages).

On Apple Silicon (darwin + arm64), this project now pins tensorflow-macos==2.15.0 and includes tensorflow-metal automatically via pyproject.toml, so no extra TensorFlow install step is required after uv sync.

Mac Silicon GPU Support (Optional Recovery)

If your environment was created before the TensorFlow platform pins were added, or if TensorFlow import fails after sync, reinstall the Mac runtime packages:

uv pip install tensorflow-macos==2.15.0 tensorflow-metal>=1.1.0

Then activate the environment:

source .venv/bin/activate

Advanced / Optional Topics

Standalone Mode (No Server)

Use the standalone script in example_usage if you explicitly want serverless local inference:

  • example_usage/decimer_standalone_no_server.py

For most users, the server mode above is simpler to operate and integrate.

Python Client Package

If you want the lightweight client wrapper in a separate environment:

pip install ./packages/decimerapi/

This is optional; direct HTTP calls work without it.

API Examples by Experience Level

  • Beginner local client usage: example_usage/decimer_server_usage_example.py
  • Direct raw API call (Docker/local server): example_usage/decimer_naked_api_eg_docker_only_example.py
  • Standalone, no server mode: example_usage/decimer_standalone_no_server.py

EMF Support (Optional)

If you need to convert EMF files (typically from Word document extraction), install Inkscape. The API will ignore EMFs otherwise.