What's the difference between last.pt&best.pt, and which one should i use when i test my custom data?

[PromiseXu1]

I have tried your work of yolov3 on my custom data, and when i train on yolov5 i find that yolov5 is more excellent! I just train for 100 epochs, the map50 have reached at 80! And i will train more epochs to see if it could be better .

But the question is that when i finish the train, and start to test, which weight.pt should i choose? And what's the difference between last.pt and best.pt?

Could you help me? Thanks sincerely.

[github-actions]

Hello @PromiseXu1, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Jupyter Notebook , Docker Image, and Google Cloud Quickstart Guide for example environments.

If this is a bug report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom model or data training question, please note Ultralytics does not provide free personal support. As a leader in vision ML and AI, we do offer professional consulting, from simple expert advice up to delivery of fully customized, end-to-end production solutions for our clients, such as:

• Cloud-based AI systems operating on hundreds of HD video streams in realtime.
• Edge AI integrated into custom iOS and Android apps for realtime 30 FPS video inference.
• Custom data training, hyperparameter evolution, and model exportation to any destination.

For more information please visit https://www.ultralytics.com.

[glenn-jocher]

@PromiseXu1 last.pt are the weights from the last epoch of training. best.pt are the best weights recorded during training.

[PromiseXu1]

@glenn-jocher ah it seems so obviously, thanks for your patient... but i find sometimes best.pt not always be there, and i find someone have mentioned it, are you dealing with it?

[PromiseXu1]

[TaoXieSZ]

@PromiseXu1 Because in exp11, when you resume from another exp, the model you trained didn't have better mAP than the one you resume from.

[glenn-jocher]

@PromiseXu1 --resuming still has a few bugs to iron out, so I would avoid it as much as possible.

Best.pt will be saved anytime a new best fitness is observed on your validation set. We define fitness as a weighted combination of metrics:

yolov5/utils/utils.py

Lines 881 to 885 in 48e15be

	def fitness(x):
	# Returns fitness (for use with results.txt or evolve.txt)
	w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
	return (x[:, :4] * w).sum(1)

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Yu-Hang]

yolov5/utils/utils.py

Lines 881 to 885 in 48e15be

	def fitness(x):
	# Returns fitness (for use with results.txt or evolve.txt)
	w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
	return (x[:, :4] * w).sum(1)

Does this mean the weights for new best fitness will be recorded even if the model already overfits?

[glenn-jocher]

@Yu-Hang they are exactly as they sound, the last and best checkpoints during your training.

If the model overfits then by definition there will be no new best fitness.

[Yu-Hang]

Let me ask more specifically. The fitness function does not include any loss metrics. So, if you train for a large epoch, [P, R, mAP@0.5, mAP@0.5:0.95] are increasing, but some val loss starts to increases as well (which indicates overfitting?). In this case, will this creates new best fitness?

[glenn-jocher]

@Yu-Hang fitness function is defined here. best.pt is saved on every new best-fitness epoch.

yolov5/utils/metrics.py

Lines 13 to 17 in 417a2f4

	def fitness(x):
	# Model fitness as a weighted combination of metrics
	w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
	return (x[:, :4] * w).sum(1)