Request to Investigate Unusually Long Training Time and Credit Usage

🤝 Community Help
1

Describe your question/issue here! (delete this when you post)

  • Project Type: instance segmentation

  • Operating System & Browser: Google chrome window

  • Project Universe Link or Workspace/Project ID:keg_pallet_hole-5-0, yamato-dnvf7/keg_pallet_hole-5-0-3-yolo26s-seg-t2

  • **Do you grant Roboflow Support permission to access your Workspace for troubleshooting? (Yes/No): yew
    **
    Unusually Long Training Time and Unexpected Credit Usage

    Hello Roboflow Support Team,

    I would like to report an issue regarding unusually long model training time and unexpected credit usage.

    I have been training YOLO26 instance segmentation models using the Previous Checkpoint option.

    My previous training job (t1) completed relatively quickly with reasonable credit usage. However, the subsequent training job (t2) took approximately 17 hours and 37 minutes, despite using the same dataset version.

    Training Information

    • Model Name: t2

    • Training Job ID: 00bbb11fe9592e6fa932

    • Dataset: Version 3

    • Total Images: 4,754

    • Training Images: 4,236 (including augmentation)

    • Expected Credits: Approximately 10 credits

    • Workspace Training Usage (October 28–29): 37.39 credits

    I understand that 37.39 credits represents the total workspace training usage and may not be entirely attributable to t2.

    Questions

    1. Why did t2 take significantly longer than t1, even though both used the same dataset version?

    2. Was the training delayed by a server-side issue, or was the entire duration spent on GPU training?

    3. Was the Previous Checkpoint option correctly applied?

    4. How many epochs were executed during t2 training?

    5. How many credits were actually charged for this specific training job?

    If the excessive training time or credit usage was caused by a platform issue, I would appreciate a review and possible credit adjustment.

    Thank you for your support.