Skip to content

CHANGELOG.md

Latest commit

 

History

History
43 lines (32 loc) · 1.62 KB

CHANGELOG.md

File metadata and controls

43 lines (32 loc) · 1.62 KB
Raw

Changelog

All notable changes to this project are documented here. The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

Unreleased

0.3.0

Added

  • BYOL objective (BYOLModel) with a scheduled EMA target network.
  • Embedder / load_pretrained high-level embedding API with mean/max/cls pooling.
  • Offline spectrogram precomputation (precompute_folder, PrecomputedMelDataset).
  • resona embed CLI subcommand and YAML config support.
  • docs/ set: architecture, usage, design notes and API reference.

Changed

  • Trainer now auto-coerces spectrograms to the encoder's expected shape.

0.2.0

Added

  • Contrastive objective (ContrastiveModel) with the NT-Xent loss.
  • Spectrogram augmentations: SpecAugment, mixup, random resize crop, noise.
  • Trainer with warmup-cosine scheduling and gradient clipping.
  • resona pretrain / resona info CLI.

Changed

  • Positional embeddings switched to fixed 2-D sin-cos.

0.1.0

Added

  • Pure-PyTorch log-mel frontend (LogMelSpectrogram) with Slaney/HTK filterbanks.
  • Patch embedding and the MAE-style random_masking / restore_tokens utilities.
  • AudioTransformer encoder and the MaskedSpectrogramModel objective.
  • Synthetic dataset, project scaffolding, CI and tests.