fixerpy 0.1.0

Fixerpy helper to clone/build/run the NVIDIA Fixer container and run inference on example images.

Usage requires Docker (with GPU support for best performance) and the uv tool.

Quickstart

• Clone + build + download weights + run inference on Fixer examples:
  • uv run fixer

Details

• What uv run fixer does:
  • Clones https://github.com/nv-tlabs/Fixer.git into .fixer_work/Fixer (if missing)
  • Builds the Docker image from Dockerfile.cosmos with tag fixer-cosmos-env
  • Downloads model weights via hf download nvidia/Fixer --local-dir models into the cloned repo
  • Runs the container per Fixer README, mounting the repo to /work and invoking the inference script on example images:
    • python /work/src/inference_pretrained_model.py --model /work/models/pretrained/pretrained_fixer.pkl --input /work/input --output /work/output --timestep 250
  • Outputs are saved to ./output (relative to current working directory)

CLI options

• --dest-root PATH: Working directory (default: .fixer_work)
• --input PATH: Input directory (default: <dest-root>/Fixer/examples)
• --output PATH: Output directory (default: ./output)
• --timestep INT: Inference timestep (default: 250)
• --batch-size INT: Batch size (default: 1)
• --test-speed: Run speed benchmark before inference
• --no-gpus: Disable GPU pass-through for the container
• --platform linux/amd64|linux/arm64: Build target platform (auto-detected by default)

Examples

• Run with defaults (examples → output):
  • uv run fixer
• Run with custom input/output:
  • uv run fixer --input /path/to/images --output /path/to/out --timestep 250
• Force platform (for cross-arch hosts or when needed):
  • uv run fixer --platform linux/amd64

Notes

• This project uses Docker and the Fixer README’s recommended /work mount layout.
• Weights are downloaded to .fixer_work/Fixer/models via Hugging Face CLI.
• Ensure your Docker supports GPU (--gpus all) if you want GPU acceleration.