Skip to content

LR for MARS #8

@lishuai-97

Description

@lishuai-97

Is there any guidance on how to tune lr and lr_1d for MARS?

I noticed that in the README, under the hyperparameters for GPT-2 models, lr is set to be 2× lr_1d. Is this ratio optimal, and does it also apply to other models?

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