aludam-fullplate

The neural tier of aludam: same import, same syntax, same DetectionResult - plus the heavy dependency stack the four detectors need to actually run.

pip install aludam-fullplate

is exactly

pip install "aludam[neural]"

aludam-fullplate is a thin wrapper whose only dependency is aludam[neural], so there is one dependency list to maintain and two names to choose from. Both install torch, transformers, huggingface_hub, timm, peft, opencv-python-headless, librosa, soundfile, scipy and onnxruntime.

Usage

Identical to aludam - either import works:

from aludam import load_detector        # or: from aludam_fullplate import ...

det1 = load_detector("text_det");  result1 = det1.predict(text)
det2 = load_detector("img_det");   result2 = det2.predict(image_path)
det3 = load_detector("vdo_det");   result3 = det3.predict(video_path)
det4 = load_detector("aud_det");   result4 = det4.predict(audio_path)

print(result1)
# DetectionResult(score=0.8012, label='ai', confidence=0.5582, explanation='...')

They are the same objects: aludam_fullplate re-exports aludam, it does not wrap it. The result class, the errors, the CLI (aludam ...) and the weight tooling all come from the one aludam package.

Weights are fetched, never bundled

aludam fetch          # download model weights into the HF cache, once
aludam status         # what this install is missing

Wheels ship code and declarations only. Missing dependencies and missing weights raise (AludamDepsMissing, AludamWeightsMissing); they never degrade into a heuristic-looking score.

Calibration - read this

runtime["calibrated"] ships False: thresholds and confidence are not fitted yet, so label is the upstream detector's own verdict and details["placeholder_label"] shows what the unfitted band would have said. See the aludam README for the full rule.

Status

Alpha. No accuracy claim is made for either tier until Phase 0 of the plan measures one; aludam-fullplate exists to carry the neural stack, not to advertise numbers. See ALUDAM_PLAN.md in the repository root.

License

MIT - see LICENSE. Third-party model weights are fetched, not bundled; see THIRD_PARTY_NOTICES (full inventory in the aludam wheel's copy).