MSST Bible

___(old) VR settings for ensemble (section deprecated, see the section here)__

Page 188 of 228 · Open the original GDoc ↗

 

Ensemble can provide different results from one current main model, but not especially better in all cases, so it’s also a matter of taste and conscious evaluation.

 

  • Aggressiveness shouldn’t be set to more than 0.1

(also check 0.01)

  • high_end_process: bypass (official recommendation) or mirroring 2 (in some cases)
  • In most cases, you shouldn’t use more than 4 models to not decrease the quality (developer recommendation)

 

Don't use postprocessing in HV Colab for ensemble (doesn't work).

 

Other recommended models for ensemble:

 

HP2-4BAND-3090_4band_arch-500m_1.pth,

HP2-4BAND-3090_4band_arch-500m_2.pth

(+new 3 band?)

as they currently the best (15.08.21) but feel free to experiment with more (I also used old MGM beta 1 and 2 with two above,

some people also used vocal models as well, and later there was also HP2-MAIN-MSB2-3BAND-3090_arch-500m model released, which gives good results solo).