Skip to content

On the top-K #155

@GAOSongAAA

Description

@GAOSongAAA

Hello, and apologies for all the questions. I'm running into an out-of-memory (OOM) error during inference because my 5070 GPU only has 12GB of VRAM. I can't upgrade my GPU right now, but I've noticed that lowering the top-k value in the Refine module significantly reduces VRAM usage.

My questions are:

  1. What does a lower top-k value mean? My intuition is that a very low top-k value will lead to a significant drop in accuracy, especially for long-chain proteins or those with complex folds (like cross-chain β-sheet interactions). Do you have any comments on this?

  2. If I set top-k to 8, will the model still be effective?

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