Trian coco128 from scratch get 0 mAP

[yxNONG]

• GPU = 1080 Ti
• Python = 3.7.7
• PyTorch = 1.5
• numpy = 1.17

thinks for your great job !
i download the coco128 and try to train it from scratch:

python train.py --img 640 --batch 16 --epochs 10 --data ./data/coco128.yaml --cfg ./models/yolov5s.yaml --weights '' --device 4，5，6，7

The GIou, objectness, Classification work, However, the mAP, Precision and Recall is both 0, in each epoch.

   5/9     11.1G    0.1098   0.08732     0.117    0.3142       200       640         0         0         0         0    0.1034   0.07643    0.1161

Then, i try to use the pretrained weight:
python train.py --img 640 --batch 16 --epochs 1 --data ./data/coco128.yaml --cfg ./models/yolov5s.yaml --weights yolov5s.pt --device 4，5，6，7

in this case, everything work well

   0/9      7.3G   0.04367   0.07351   0.01972    0.1369       177       640    0.4245    0.7103     0.657    0.4169   0.04242   0.05659   0.01382

i get this with the original code.
since i didn't implement the Nvidia Apex, i change the mixed precision training = False
simply run again and get the same result.( mAP = 0 )

any idea about this ? thanks for your reply.

[github-actions]

Hello @yxNONG, 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 that 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.

[yxNONG]

i try 50 epoch, get the mAP@50 great then 0 after 15 epoch, the problem is simply due to that train from scratch need more time to converge.

[glenn-jocher]

@yxNONG yes you are correct, training from scratch takes longer to converge. Note that when training extremely small datasets like coco128.yaml from scratch, minimum training is 300 epochs, and recommended training is up to 1000 epochs.