Thanks to visit codestin.com
Credit goes to github.com

Skip to content

after about 1000 training steps, training loss continues to rise #88

@Alex417ii

Description

@Alex417ii

When training the AnySplat model on the DL3DV_10k_samples dataset, I kept most configurations unchanged except for setting num_context_views: 8, max_img_per_gpu: 4, and the data loading thread number to 1. However, after about 1000 training steps, I encountered the message: "Skipping batch with high loss (0.249768) at step 1001 on Rank 25". Subsequently, the training loss continues to rise, even reaching around 1.5. Could you explain what might be causing this and suggest any possible solutions? I look forward to your reply. Thank you!

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