Set a seed of generator with an option for more randomness when training several models with different seeds

[developer0hye]

It is related to #9545.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancing reproducibility of YOLOv5 training runs with a new seeding option.

📊 Key Changes

• Added a seed parameter to data loader functions to set the random seed for shuffling data.
• Updated the random seed generation logic in create_dataloader by incorporating the new seed value.
• Modified the train.py script to pass the seed parameter from the command-line options to the data loaders.

🎯 Purpose & Impact

• Reproducibility: The addition of a seed parameter allows for more reproducible training runs, as the data shuffle order can be kept consistent across different runs, given the same seed value.
• Customization: Users have more control over their training process, as they can set the seed manually if desired.
• Flexibility: This change is backward compatible, meaning it will not affect existing workflows unless users opt to set their own seed.
• Consistency: Data loading, especially when shuffling, can now be consistent across distributed training setups with the same seed, aiding in debugging and result verification.

[glenn-jocher]

@AyushExel do you think we should change dataset shuffling on --seed changes? Right now I don't think this is explicitly enforced, so different --seed values will likely still see the same dataset generator seeds.

[AyushExel]

@glenn-jocher yes we should if the deterministic behaviour is still preserved

[glenn-jocher]

@developer0hye PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐