--notest bug fix

[NanoCode012]

Command to replicate:

python train.py --weights yolov5s.pt --epochs 3 --img 320 --device 0 --notest

Mentioned in Issue #475 by @cesarandreslopez

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Optimization of model saving conditions during training in YOLOv5.

📊 Key Changes

• Removed indentation, changing the scope of model evaluation and saving logic.
• Adjusted the conditional logic for saving training results and updating the best model performance.

🎯 Purpose & Impact

• Ensures consistent and accurate saving of the model's performance data after each training epoch.
• Improves clarity of the training process, potentially leading to more reliable model saving and performance tracking for developers and users.
• No new features or performance improvements, but a cleaner codebase can lead to fewer bugs and more straightforward maintenance.

[glenn-jocher]

@NanoCode012 --notest is supposed to record all normal metrics even if it doesn't test. An empty results file is not the intended result.

[NanoCode012]

Ah, I see. Thanks for the fix!

[glenn-jocher]

It looks like some commands were accidentally dropped an additional level into an if statement, so I've simply removed them back where they should be. I tested in the colab notebook and results.txt looks all good (attached).
results.txt

[glenn-jocher]

Looks like tests are passing. I'll merge. This should fix https://docs.ultralytics.com/yolov5/tutorials/multi_gpu_training.