PyraFuse v1.1.1

PyraFuse v1.1.1

Feature release adding a reproducible training workflow for skin, fabric, and
background semantic segmentation.

Highlights

• Added scripts/train_segmenter.py training CLI.
• Added configurable encoder, decoder, loss, augmentation, split, EMA, and
  checkpoint options.
• Added --smoke and --dry-run validation modes.
• Added automatic class-weight estimation and metric reporting.
• Added CPU/GPU training examples to the README.
• Improved CUDA environment setup and safe device defaults.
• Added stronger CLI argument validation.
• Updated package version to 1.1.0.

Validation

• PyPI wheel and source distribution built successfully.
• pytest: 3 passed.
• One-batch CPU training smoke test completed successfully.
• CI passed on Python 3.12 and 3.13.

Installation

pip install pyrafuse==1.1.0

PyraFuse v1.0.0

PyraFuse v1.0.0

Initial public release of PyraFuse, a DINOv3-based semantic segmentation
framework for skin, fabric, and background detection.

Included

• PyTorch inference and model-zoo API
• Small, Small+, Base, and Large pretrained models
• Hugging Face model repository integration
• ONNX/TensorRT export support
• INT8 calibration and verification workflow
• Accepted AIMLSystems 2026 manuscript
• Apache-2.0 source-code license

Install from PyPI after publication:

pip install pyrafuse