Update train.py

[glenn-jocher]

• return hyp.scratch.yaml as default for both from scratch and finetuning training Add InfiniteDataLoader class #876 (comment)
• simplified ckpt names to avoid appending run name to weights
• force replot of test_batch0_gt and test_batch0_pred on final epoch

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced cleanup and default parameter settings in YOLOv5 training script.

📊 Key Changes

• Added glob import to use globbing patterns for file operations.
• Automatically remove prediction images before final evaluation to declutter the output directory.
• Simplified the generation of result file names by improving the naming logic.
• Set a default value for the --hyp argument to 'data/hyp.scratch.yaml', standardizing hyperparameter selection.
• Removed redundant conditional check for assigning the opt.hyp argument.

🎯 Purpose & Impact

• 🧹 Cleanup: Removing prediction images ensures a cleaner workspace, which helps users navigate results without confusion from outdated files.
• 🏷️ Naming Consistency: Simplifying result file names reduces complexity and potential confusion when interpreting output files.
• 📈 Usability: By having a default hyperparameter path, new users can start training models with sensible defaults without specifying additional arguments.
• 🔍 Clarity: By cleaning up the code and removing unnecessary checks, the script becomes more maintainable and easier to understand, benefiting all users.

[glenn-jocher]

All updates look good, passing CI, will merge.

These are small tweaks to improve training results, visualization and usability.

[NanoCode012]

Hi @glenn-jocher , may I ask why you would only append n to best/last/results if opt.names is numeric?

I think I found one bug in code with the renaming.

yolov5/train.py

Lines 366 to 368 in 281d78c

	fresults, flast, fbest = 'results%s.txt' % n, wdir / f'last{n}.pt', wdir / f'best{n}.pt'
	for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', 'results.txt'], [flast, fbest, fresults]):
	if os.path.exists(f1):

results.txt is not being replaced.

My proposed solution

fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt' 
for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file], [flast, fbest, fresults]): 

Result from my proposed fix with python train.py --name 2:

[glenn-jocher]

@NanoCode012 ah, yes that change works better than what I have, not sure why I was treating results.txt differently. Can you do a quick PR?

The isnumeric thing is a bit of a personal hack. I'm currently keeping historical records of experiments in GCP buckets with numeric names (i.e. last241.pt, last242.pt etc), but I'm only sending the 3 files here rather than the entire runs/exp directory so I need to keep the numeric names attached for now. In the future I'm planning on switching to full log_dir saving, and then I can delete the isnumeric line, and then the weights and results should not have any name appended to them (for easier plotting and testing).

[NanoCode012]

yes that change works better than what I have, not sure why I was treating results.txt differently. Can you do a quick PR?

Sure thing!