Skip to content

Training non-deterministic: which seed reproduces pretrained results? #140

Description

@ShenZheng2000

Hi, thanks for the great work!

I retrained the model twice using the same data and code on the DanceTrack val set. One run reproduces the pretrained results, while the other shows noticeable differences in IDF1 and IDs (see screenshot).

Image

While checking the code, I noticed that the default seed is None, which makes training stochastic:

parser.add_argument("--seed", default=None, type=int, help="eval seed")

Could this be the reason for the variation?
If so, which seed was used to obtain the pretrained results?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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