Skip to content

ViT web12m training recipe and loss log/final loss value #40

@adit-0503

Description

@adit-0503

Hi @mk-minchul could you describe the training recipe of ViT-Kprpe-web12m - I am training using the recipe given in the paper i.e lr = 1e-4 and batch_size of 1024 ( using ddp over two gpus) + the gridsample config file with adamw and momentum and decay as given in the supplementary of the paper with pfc=0.3 (since it is faster) - but my results are not as good. Could you confirm whether this is the same recipe for your pretrained model or is it just my partial_fc value which is causing the issue?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions