Codename-RVC-Fork-V3.0.3-rev2

Release of the version: 3.0.3-rev2

Notes:

• Latest changes from the main applio repo ( including refinegan changes )
• Complete yeeting of half precision training ( fp16 )
• Minor improvements and misc

Codename-RVC-Fork-V3.0.3-rev1

Release of the version: 3.0.3-rev1

Notes:

• More checks for stuff at init ( Benchmarking, deterministic etc. )
• Better multi-gpu-scenario gpus / devices assignment.
• Various optimizations incl. more in-place operations

Codename-RVC-Fork-V3.0.3

Release of the version: 3.0.3

Notes:

• New logging mechanism for losses: Average loss per epoch logged as the standard loss,
  and rolling average loss over 5 epochs to evaluate general trends and the model's performance over time.
  Both work for each metric individually.

• Features a different optimizer: RAdam ( Rectified Adam )
  ( More init. stability and compared to AdamW, doesn't require using / configuring warmup. )
  Most likely better convergence / generalization on average, compared to plain AdamW without a warmup.

• Few tweaks for the ui, some formatting changes and generally, updated in-line with Applio main repo.

PS. Do not turn on warmup. RAdam handles all of that on it's own. It's kept in for the sake of future AdamW's optional usage.

Codename-RVC-Fork-V3.0.2

Release of the version: 3.0.2

Notes:

• Both Generator and Discriminator now fully support checkpointing and in-place operations.
• Some precision related changes for inference
• Various optimizations and UI related changes.
• New options for preprocessing ( including configuration in relation to 'mute files' usage )

Codename-RVC-Fork-V3.0.1-rev1

Release of the version: 3.0.1 - revision 1

Notes:

• New Envelope loss function added
• Checkpointing added
• Various optimizations
• Some minor logging related tweaks ( i.e. improved prints' readability and such. )

Rev1 notes:

• Corrected ' mel_processing '

v3.0.0-rev1: 🐛fix

Revision: 1 of the 'v3.0.0' release.

Notes:

• 🐛fix: A bug fix that fixes an issue of training not working when user doesn't doesn't change the n steps value from 0 to positives ( or if decides to not use the metric )
• Better description for averaging metric