3.3.0

What's Changed

• Weights now download from the Hugging Face Hub 🤗 instead of a GitHub release asset: https://huggingface.co/phamquiluan/ResidualMaskingNetwork
  • Checkpoints are cached in ~/.cache/huggingface via hf_hub_download; pre-downloaded local files (pretrained_ckpt, face_detection_yunet_2023mar.onnx in the working directory) still take precedence.
  • Dependency change: tqdm removed, huggingface_hub added.
• All FER2013 benchmark checkpoints are also mirrored on the HF Hub under benchmarks/.
• Paper page: https://huggingface.co/papers/2603.05937

Full Changelog: 3.2.0...3.3.0

3.2.0

What's Changed

• Replace Caffe SSD face detector with YuNet (cv2.FaceDetectorYN) for OpenCV 5 compatibility (#61)
• Require opencv-python>=4.8.0 (works on both OpenCV 4.x and 5.x)

Fixes AttributeError: module 'cv2.dnn' has no attribute 'readNetFromCaffe' on fresh installs, caused by OpenCV 5.0 removing the legacy Caffe importer from the dnn module. The face detection output format and public API are unchanged.

Full Changelog: 3.1.2...3.2.0
PyPI: https://pypi.org/project/rmn/3.2.0/

Fixed #40

v3.0.4

Fixed #40

Chores

• Update URL for pre-trained weights, thanks to dmitry.pigasin@yandex.ru

v3.0.0

v3.0.0

v3.0.0a5

v3.0.0a5

v2.0.0

v2

v1.0.1

Bump version to v1.0.1

v0.0.6

Bump version

v0.0.2

Update version