8000 Training Degrades After ~3000 Steps on NeRSemble Multi-View Dataset · Issue #8 · gapszju/RGBAvatar · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content
Training Degrades After ~3000 Steps on NeRSemble Multi-View Dataset #8
New issue
New issue
Open
Open
Training Degrades After ~3000 Steps on NeRSemble Multi-View Dataset#8
@KEH0T0

Description

@KEH0T0
KEH0T0
opened on May 9, 2025

I switched the code to use NeRSemble’s multi-view dataset and ran training. Since it contains around 5,000 images, perhaps due to the large data volume, training with train_offline seems to proceed normally up to about 2,000 steps, but then the results begin to deteriorate from around 3,000 steps onward. Could you share your thoughts on whether this is a bug in the code or a hyperparameter mismatch? I’ve attached the output screenshots/results below.

Image
Image

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
      0