v0.1.2

• We kept backward compatibility with previous version for everything!
• Added configurable cache-based weight handling across fovea-od-localization, quality, and registration inference
  • Added centralized download and extraction utility with retries, safer extraction, and rollback cleanup for partial extraction/download artifacts
  • Improved error handling/messages for download failures and manual workaround guidance.
  • This also fixes #29
  • From now on, pass cache_dir argument or set FIT_CACHE_DIR to use a non-default directory to store pretrained weights at. Backward-compatibility was ensured.
• quality_prediction
  • make the image size, that the images are resized to internally, configurable. Kept backward-compatible behavior by setting default to 512 (as in <= v0.1.1).
• registration
  • Fixed original code to handle modern torch versions: align_corners argument was wrong if torch version was >= 2.
• vessel segmentation:
  • Fixed dtype argument deprecation in newer numpy versions.
• circle_crop
  • Failure handling is now informative and non-fatal, e.g. when applying inside dataset loops: failures now emit UserWarning, return shape-preserving zero outputs, and mark failure with radius=-1.
  • Fixed the process_img uninitialized-variable bug by ensuring deterministic fallback failure outputs after exceptions. This fixes #28
  • Changed crop sizing behavior to square-only (size int), with backward-compatible tuple/list support using the first value.
• utils
  • ImageTorchUtils:
    • Added robust greyscale image support (in to_tensor, to_batch, squeeze)
    • Added handling for more dtype and shape combinations
    • Fixed numpy 2 deprecated np.array() calls
  • basics:
    • added greyscale support in show(), thanks to @Page0526
• Improved printed messages, e.g. when downloading weights.
• dev: added tox-uv for unittests