Sam3 Training Failure

Hello,

Recently tried to train a SAM3 model but got some unexpected behavior during training.

Details:

  • Instance segmentation project
  • MacOS, running Brave browser
  • Project: Sign in to Roboflow
  • Yes, I grant support permission to help me troubleshoot

Around epoch 28 of training, the mAP dropped from ~50% to 0%, and the box / object losses skyrocketed. See screenshot:

Hi @wharper,
My name is Ford and I am a Support Engineer at Roboflow. Happy to help here!

I am currently diving in on our end, and will keep you apprised of my progress.

Thank you for bringing this forward!