FROM nvcr.io/nvidia/pytorch:22.04-py3

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

This PR updates dependencies and handling of TensorRT for better compatibility and cleanliness in Docker environments.

📊 Key Changes

• Improved exception handling for importing TensorRT in export.py, falling back to requirements check if the initial import fails.
• Updated the NVIDIA PyTorch Docker base image to version 22.04-py3.
• Removed a large unnecessary directory (/opt/pytorch) in the Docker image to save space.
• Changed the location where Arial fonts are downloaded in Dockerfiles for consistency and potential reduction in layer sizes.
• Removed the installation of specific versions of PyTorch and torchvision, replacing it with a command to install the latest version that matches the CUDA 11.3 version, and enforced the Pillow version to be >=9.1.0.

🎯 Purpose & Impact

• The changes in handling TensorRT import issues enhance robustness, ensuring the availability of necessary packages for GPU support without manual intervention.
• The Docker image update reflects a move towards more current dependencies, promoting compatibility with recent codebases.
• Space optimization in Docker images through removing unnecessary directories will lead to quicker download and deployment times for users.
• Streamlining the font downloads into a common layer in Dockerfiles may improve build efficiency and reduce image sizes.
• By allowing automatic installation of compatible PyTorch and torchvision versions, the project avoids potential conflicts with pre-existing installations, while specifying the Pillow version ensures compatibility with other requirements.