[Python] Jupyter Notebook and Python Script return different YOLOV8 instance segmentation masks

I've been running a pretrained YOLOV8 model for instance segmentation. For testing and development I've been using an Anaconda Jupyter Notebook, which I converted into a Python script before deployment.

Here is the (correct) mask from the Jupyter Notebook:



and here is the result from the Python script:



You can see that the Python result has weird blocks towards the edges of my head and my phone.

• The code is exactly the same for both versions
• The Python virtual environment is directly created from the Jupyter environment export (and I double checked: identical Python version and identical versions for all libraries)
• Same CUDA version
• Same hardware (both examples were executed on the same machine)

Does anyone know why this is happening or how I can fix it?

Continue reading...