cudnn.benchmark = True on Seed 0

[glenn-jocher]

Signed-off-by: Glenn Jocher glenn.jocher@ultralytics.com

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved the random number generator initialization for potentially faster training.

📊 Key Changes

• Removed unnecessary imports and settings for cuDNN when initializing seeds.
• Set torch.backends.cudnn.benchmark to True by default for all training sessions.
• Confined deterministic settings within a conditional block that's only executed if determinism is requested.

🎯 Purpose & Impact

• The removal of some cuDNN settings streamlines the seed initialization process, making the code more readable and potentially reducing overhead.
• Enabling cudnn.benchmark by default aims to speed up training by allowing cuDNN to automatically find the most efficient algorithms for the hardware.
• The careful placement of deterministic settings preserves reproducible results when needed, without affecting performance otherwise.
• Users should see improved training speeds with no change in usage, while still having the option for deterministic behavior when reproducibility is crucial.

[glenn-jocher]

@AyushExel enabled cudnn.benchmark=True by default (affects all tasks). Still deterministic, but cuts COCO training time a bit.

# Master
      Epoch    GPU_mem   box_loss   obj_loss   cls_loss  Instances       Size
        2/2      14.7G    0.03588    0.05798    0.01502        157        640: 100%|██████████| 7393/7393 [17:42<00:00,  6.96it/s]                                   
                 Class     Images  Instances          P          R     mAP@.5 mAP@.5:.95: 100%|██████████| 157/157 [00:32<00:00,  4.82it/s]                          
                   all       5000      36335      0.713        0.6      0.649      0.466

3 epochs completed in 0.964 hours.


# PR
      Epoch    GPU_mem   box_loss   obj_loss   cls_loss  Instances       Size
        2/2      9.23G    0.03588    0.05798    0.01502        157        640: 100%|██████████| 7393/7393 [16:50<00:00,  7.31it/s]                                   
                 Class     Images  Instances          P          R     mAP@.5 mAP@.5:.95: 100%|██████████| 157/157 [00:31<00:00,  5.06it/s]                          
                   all       5000      36335      0.713        0.6      0.649      0.466

3 epochs completed in 0.931 hours.