Skip to content

How to obtain Ground Truth embeddings for evaluation reproduction? #1

@Rankyer

Description

@Rankyer

Thank you for the impressive work on LLM2Fx!

I am currently trying to reproduce the evaluation results (MMD scores) as described in the paper, usingllm2fx/evaluation/text2fx/llm_eval.pyand upperbound.py.

I noticed that these scripts rely on Ground Truth (GT) embeddings, which are currently pointing to a hardcoded local path in the code:

  • In upperbound.py: EMB_DIR = "/data3/seungheon/fx_embedding_new"

  • In llm_eval.py: It requires the argument args.gt_embeds_dir.

Could you please clarify the standard workflow to obtain these GT embedding files?

Are these pre-computed embeddings available for direct download? Or is the intended method to manually re-synthesize the SocialFX audio using the fx-processor repository and then run llm2fx/evaluation/embeds/fx_encoder.py to generate the .npy files locally?

Any guidance on how to properly set up the gt_embeds_dir to run the evaluation scripts would be greatly appreciated.

Thanks!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions