Skip to content

Strange phenomenon in VTAB-1k #3

Description

@Tommy-yours

Hi, nice work! When I reproduce your work and other PEFT methods, I find that methods like NOAH\SSF use VTAB data in format of picture which is converted by NOAH code: https://github.com/ZhangYuanhan-AI/NOAH/blob/main/data/vtab-source/get_vtab1k.py. However, in VPT and your work, data processing comes from original VTAB (data maybe tensor format). I wonder do these two forms (picture and tensor) of data have any impact on training? Here are some strange phenomenon:
(1) cifar100: models trained by picture format always achieve ~70% acc (like NOAH\SSF), while models trained by tensor format always achieve ~80% acc. In general, the same model trained by different format data of cifar100, will obtain results with relatively large differences.
(2) LoRA methods: in NOAH (https://arxiv.org/pdf/2206.04673) and your work, LoRA performance on cifar100\svhn\sun397... have a big difference.
This problem has troubled me for a long time, I guess there are several possibilities:
(1) dataset version: data in NOAH are converted by a different version.
(2) The converted pictures maybe are not aligned with tensor format data.
Do you have any ideas, looking forward to your reply !!!

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