Skip to content

Inconsistent Results for MobileVIT v2-0.5 Models #111

@tapeimrchuang

Description

@tapeimrchuang

Dear Developers,

I trained MobileViT v2-0.5 using the latest codebase and obtained results that do not match the numbers reported in the original paper. Specifically, on the ImageNet‑1K dataset, my run achieved a Top‑1 accuracy of 68.74%, whereas the paper—and the released pretrained weights—report 70.17%. I have not modified the model or configuration files, and I followed the official training procedure. Could you let me know if any recent changes to the MobileViT v2-0.5 code might explain this discrepancy?

My hardware configuration is as follows:
GPU: 2 × RTX 4090

I would like to ask:
Will hardware differences cause variations in model accuracy?
Additionally, if I follow your instructions exactly to train from scratch on ImageNet-1K, will I obtain the same accuracy reported in the paper? Or should I expect some deviation? If so, how much deviation is typical?

Thank you for your time and assistance on this matter. I look forward to your reply.

Best regards,
Sean

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