Skip to content

Jabberjay v0.1.2

Choose a tag to compare

@github-actions github-actions released this 26 Oct 00:53
· 1 commit to main since this release
Immutable release. Only release title and notes can be modified.
26b0cec

Fixed

  • SpectraAASIST and SpectraAASIST3 predictions were effectively random β€”
    lab260's checkpoints nest each AASIST encoder residual block one level
    deeper (aasist.enc.enc.N.0.*) than our port (aasist.enc.enc.N.*).
    PyTorchModelHubMixin.from_pretrained() defaults to strict=False, so the
    mismatch was silent and the entire six-block encoder stayed randomly
    initialised β€” scores changed from process to process and the bundled
    sample clips were misclassified. Each block is now wrapped in its own
    nn.Sequential to match upstream, and logits are verified bit-identical
    to lab260's reference model.py for all three Spectra models

Changed

  • Spectra models now load strictly β€” load_pretrained() passes
    strict=True, so any future checkpoint/architecture drift raises instead
    of silently leaving layers uninitialised
  • Spectra first-run download halved β€” the Wav2Vec2-XLS-R-300M encoder is
    now built from its config alone instead of downloading the 1.27 GB
    facebook/wav2vec2-xls-r-300m weights, which every Spectra checkpoint
    overwrites with its own fine-tuned copy anyway. Also removes the
    Wav2Vec2Model LOAD REPORT noise printed on every Spectra load
  • VIT spectrograms render without pyplot β€” get_image() draws onto a
    bare matplotlib.figure.Figure, so no GUI backend is started. Output is
    pixel-identical, but detect() with VIT now works off the main thread and
    on Python builds without Tk
  • ty invalid-mutable-override ignored under Models/ β€” ty 0.0.89
    flags every idiomatic nn.Module.forward() override because torch types
    forward as a mutable Callable attribute

Dependencies

  • huggingface-hub 1.31.0 β†’ 2.2.0, transformers 5.17.0 β†’ 5.20.1,
    torch 2.14.0 β†’ 2.14.1, pillow 12.3.0 β†’ 13.0.0, numpy 2.5.3 β†’ 2.5.4,
    scikit-learn 1.9.0 β†’ 1.9.1, matplotlib 3.11.1 β†’ 3.11.2 (lockfile;
    verified with the full model sweep β€” every non-Spectra result is
    identical to v0.1.1)
  • Dev: black 26.10.1, ruff 0.17.1, ty 0.0.89, pre-commit 4.7.0
    (pre-commit hook revs bumped to match); astral-sh/setup-uv 10.3.0