blinklinmult

BlinkLinMulT: transformer-based eye blink detection.

Version 2 is a full rewrite of the published 1.x code. It keeps the two tasks of the paper — eye state recognition (is this eye closed?) and blink presence detection (does this window contain a blink?) — and replaces everything around them:

• blinklinmult.preprocess turns each raw corpus into a uniform data/processed/<db> tree, one dataset at a time.

• blinklinmult.data collects those trees into per-dataset HDF5 files in the layout OmniLoader reads, and declares each dataset’s OmniLoader schema.

• blinklinmult.train trains the current LinMulT over those datasets jointly, driven entirely by YAML.

The 1.x public API (DenseNet121, BlinkLinT, BlinkLinMulT importable from blinklinmult.models, with weights auto-downloaded on construction) is gone; see docs/migration.md.

Submodules

• blinklinmult.assets
• blinklinmult.bench
• blinklinmult.compare
• blinklinmult.data
• blinklinmult.detector
• blinklinmult.fit
• blinklinmult.live
• blinklinmult.models
• blinklinmult.overlay
• blinklinmult.paper
• blinklinmult.pipeline
• blinklinmult.registry
• blinklinmult.stream
• blinklinmult.train

Attributes

__version__
PROJECT_ROOT	Repository root. Every path in a config is resolved relative to this.
WEIGHTS_DIR	Where from_pretrained caches downloaded weights.

Package Contents

blinklinmult.__version__

blinklinmult.PROJECT_ROOT

Repository root. Every path in a config is resolved relative to this.

blinklinmult.WEIGHTS_DIR

Where from_pretrained caches downloaded weights.

Passed as huggingface_hub.hf_hub_download’s cache_dir, so asking for the same model twice reuses the local copy instead of re-downloading. Under the user’s home rather than the repository, so an installed package works the same way as a checkout.