[YOLOv8.0.196] AttributeError: module 'numpy' has no attribute 'dtypes' – even with compatible numpy version

Hi, I’m having a persistent issue running YOLOv8.0.196 segmentation training (with custom hyperparameters) in Google Colab.
I keep encountering this error:

AttributeError: module 'numpy' has no attribute 'dtypes'

This happens during model training, after all installations and version checks.

I’ve tried all common solutions:

  • Downgraded/upgraded numpy (tried 1.23.5, 1.24.3)
  • torch==2.0.1, ultralytics==8.0.196, tensorboard==2.12.3, protobuf==3.20.3
  • Disabled WANDB, uninstalled jax/jaxlib, restarted Colab runtime, etc.
  • Checked that import and version checks work before training (no errors)
  • Installed all supporting libraries: opencv-python, pandas, pillow, pyyaml, etc.

Still, the error persists only at training start!

I suspect an underlying version incompatibility (maybe numpy/tensorboard/protobuf), but cannot find any requirements.txt or conda environment known to work for this YOLO/Ultralytics release.

  • Has anyone managed to run YOLOv8.0.196 segmentation in Colab recently?
  • Can you share a full working requirements.txt or point to the official stack?
  • Any workaround for the 'dtypes' numpy error with this YOLO version?

Thanks for your help!


Project Type: Segmentation

Hi @DomenicoAndreucci!
I am sorry you have run into this issue! Unfortunately I am unable to support training runs outside of the Roboflow platform.

There are a lot of uncontrollable variables outside of our platform, so we can only provide support for training runs conducted on the Roboflow platform. I apologize for the inconvenience.