Skip to content

changelog.md

Latest commit

 

History

History
124 lines (96 loc) · 4.12 KB

changelog.md

File metadata and controls

124 lines (96 loc) · 4.12 KB

Changelog

0.3.41

  • improve handing of tagged phonemes during conversion
  • removed config edits that were already added upstream

0.3.40

  • fix hidden size in OU configs
  • langloader loads on top!

0.3.39

  • dummy box has become the stretch embed
  • adds CUDA 12.9 support
  • caps Torch to 2.8

0.3.38

  • fix use_lang_id in dsdur/dspitch/dsvariance when using multiple languages but NOT merging phonemes
  • force alphabetical order in loading speakers
  • language list matches premade base dictionaries instead of localizations

0.3.37

0.3.36

  • move torch in env B back to setup script instead of directly in .yml to force the CPU version
  • support "in between" CUDA versions(versions not officially supported by Pytorch but between supported versions, rounds down to the nearest supported version)
  • force creation of onnx folder

0.3.34

  • new technically optional but recommended base env for Linux users(assets/linuxbaseenv.yml, creates difftrainerBase)

0.3.33

  • fix extra_phonemes creating a fake phoneme ['']

0.3.32

  • muon_lynxnet2
    • redirect DiffSinger download
    • change alt. backbone toggle to dropdown selection(defaults to lynxnet2)
    • using lynxnet2 requires updating tools, can be used on older versions if lynxnet/wavenet is manually selected

0.3.31

  • better main path(possibly fixing "amnesia" bug)
  • conda path fix for Linux
  • support CUDA 12.8, try 12.8 if higher

0.3.30

  • move use_note_rest to correct stage in advanced OU export

0.3.29

  • fix reloading datasets during configuration

0.3.28

  • add custom error when selecting samples so I don't have to explain ValueError every day

0.3.27

  • fix OU exports only using first speaker for pitch

0.3.26

  • multidict -> main

0.3.25

  • added CUDA 12.4 to detection in setup
  • redirect DiffSinger download to v2-backport backup

0.3.24

  • moved set config edits to setup rather than configuration
    • now if you edit things like smooth_width in the default config files, it won't be overwritten the next time you save that config

0.3.23

  • redirect DiffSinger download to main branch
  • download pc-nsf-hifigan in addition to previous nsf-hifigan
  • split breathiness/energy toggle
  • added kitchen sink config

0.3.18

  • automatic update will update core dependencies if needed
  • configs come out in the same order they went in

0.3.17-ish

  • QuickInference overhaul

0.3.14

  • added config strings to match DiffSinger update(requires running update tools)
  • updated langloader window
    • still uses langloader/merged.yaml, just has a nicer editor
    • it still usually pops up hidden behind the main window, sorry

0.3.13

  • sorry I always forget to update this section
  • there's probably a few changes that should be mentioned but I forgot
  • should actually work on Linux/Mac now
  • automates merging speakers in spk_map.json

0.3.0(multidict beta)

  • new config format for multidict setup
  • new settings files: langloader.yaml and merged.yaml
    • langloader: fixed file name/location. editable directly in DiffTrainer. lists dictionary files and global phonemes
    • merged: flexible file name/location(specified in langloader.yaml). lists groups of phonemes to merge

0.2.11

  • automatic update will update dependencies if needed
  • configs come out in the same order they went in

0.2.10

  • automatic update will require editing the old version number to 0.2.0(sorry, won't be necessary again)
  • revised basic OU export

0.2.4

  • added backbone toggle(requires Diffsinger update 11/16/24)

0.2.1

  • automated environment activation(environment names are now hardcoded, all users must use conda)

0.2.0

  • started official changelog
  • switch all users to conda or self-management
  • implement split environments
  • implement SOME for pitch estimation
  • revert to main fork of Uta's converter
  • CONVERTER: all labels MUST begin and end with [SP] (sorry for the inconvenience)
  • switch default diff_accelerator to unipc
  • implement advanced export(buggy)