FROM nvcr.io/nvidia/pytorch:22.06-py3

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Updating the base Docker image and refining Python package installations for YOLOv5.

📊 Key Changes

• 🐳 Docker base image updated from NVIDIA PyTorch 22.05-py3 to 22.06-py3.
• 🧹 Removal of a large unneeded directory (/opt/pytorch) to save space.
• 🔄 Changed the pip uninstall command to keep torch and torchvision, but remove Pillow.
• 📦 Added specific version constraints to Pillow and limited opencv-python version.

🎯 Purpose & Impact

• 🚀 Ensures compatibility with the latest software stacks by updating to a newer NVIDIA PyTorch image.
• 🛠️ Streamlines the Docker image by removing unnecessary components, which might lead to a smaller image size and reduced clutter.
• 🐍 Prevents potential conflicts by uninstalling and then reinstalling Pillow with a specific version, ensuring compatibility with other packages.
• ✨ Users get a more optimized environment for running YOLOv5, potentially improving both development and inference performance.

[AyushExel]

@glenn-jocher hey, whats the scope of this PR? I have a training task running on a docker image from 2 days ago.. I created another image and ran the same task but it's throwing a pytorch error. It works on my non gpu machine so I guess this change affected that? Here's the error:

  File "/usr/src/yolov5/utils/seg_loss.py", line 305, in build_targets_for_masks

    indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices


RuntimeError: result type Float can't be cast to the desired output type long int

Reverting pytorch to v 1.11.0 fixes it

[glenn-jocher]

@AyushExel it looks like your code is out of date a bit. You need to apply the changes in this PR from a few weeks ago to your branch and this will fix the above error.
https://github.com/ultralytics/yolov5/pull/8067/files