Skip to content

How to train a model that can fully extract the 44100hz frequency #35

Description

@dingjibang

I want to train a 2 stems model

I noticed that in the yaml configuration of each model, there are some parameters that will affect the final frequency cutoff, it seems that multigpu_drums.yaml can handle the full 44100hz frequency, but with the reduction of num_blocks (11 => 9), the model size will also decrease accordingly (29mb => 21mb).

Although using something like multigpu_drums.yaml can handle 44100hz in full, but the model shrinks instead. Does this affect the final accuracy?

It seems that dim_t, hop_length, overlap, num_blocks these parameters have a wonderful complementarity that I cannot understand, maybe this 'complementarity' is designed for the competition(mix to demucs), but I want to apply this to the real world without demucs(only mdx-net, after some testing, I think the potential of mdx-net is higher than demucs).

When I try to change num_blocks from 9 to 11, the results of inference have overlapping and broken voices... do you have any good parameters recommendations for me to train a full 44100hz one without loss of accuracy (i.e. the model does not Shrinking)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions