pretrain weights

[ou525]

When I train my own data set, do I still use the provided complete model as pre-training weights?

[github-actions]

Hello @ou525, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Google Colab Notebook, Docker Image, and GCP 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 surveillance 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]

@ou525 if your dataset is small, then yes you are strongly recommended to start training from the pretrained checkpoints:

python train.py --cfg yolov5s.yaml --weights yolov5s.pt

[ou525]

@glenn-jocher，thank you very much for your work and reply！
There is another problem, I set the training size to 640, the saved train_batch*.jpg is 1280X1280, the boxes are normal, but the test_batch*.jpg is 1280X448, and the boxes in gt.jpg are wrong

[ou525]

My dataset has only four categories, which is different from the coco dataset. So can pre-training weights be used? When I training, the error is not compatible with yolov5l.yaml, specify --weights'' or specify a --cfg compatible with yolov5l.pt

[yxNONG]

@glenn-jocher the pretrained weight is whether for the Darknet or for the whole yolo network?
if it's for the whole yolo, when i change the num of class, the yolo detect head's shape is supposed to change
in that case, can i still load the pretrained weight to the adjust yolo ?
looking for your reply, thanks

[glenn-jocher]

@ou525 @yxNONG you can request pretrained weights with train.py such as:
python train.py --cfg c --weights yolov5s.pt

It does not matter how many output classes are specified in yolov5s.yaml. All pretrained layers with matching sizes are loaded.

If you train an 80 class model, then all pretrained layers are loaded. If you train a model with different class counts, then pretrained output layers will not be loaded (because they are not the same shape). All other pretrained layers will be loaded.

For more information on this you can see our PyTorch Hub post:
https://docs.ultralytics.com/yolov5/tutorials/pytorch_hub_model_loading

[ou525]

@glenn-jocher when I updated the code to No. 16, I started training with pre-training weights, but there was a phenomenon of map 0. Is this normal?
Epoch gpu_mem GIoU obj cls total targets img_size
0/49 6.53G 0.05633 0.03247 0.0106 0.09941 17 608: 100%|████████████████████████████████████████████████████████████████████████████████| 1359/1359 [2:10:08<00:00, 5.75s/it]
Class Images Targets P R mAP@.5 mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████| 340/340 [12:07<00:00, 2.14s/it]
all 5.43e+03 2.41e+04 0.494 0.787 0.752 0.406

 Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
  1/49     6.73G   0.04195   0.02837  0.002496   0.07282        19       608: 100%|████████████████████████████████████████████████████████████████████████████████| 1359/1359 [2:14:47<00:00,  5.95s/it]
           Class      Images     Targets           P           R      mAP@.5  mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████| 340/340 [25:43<00:00,  4.54s/it]
             all    5.43e+03    2.41e+04           0           0           0           0

 Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
  2/49     6.73G   0.04216   0.03074  0.002763   0.07566        16       608: 100%|████████████████████████████████████████████████████████████████████████████████| 1359/1359 [2:18:39<00:00,  6.12s/it]
           Class      Images     Targets           P           R      mAP@.5  mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████| 340/340 [25:51<00:00,  4.56s/it]
             all    5.43e+03    2.41e+04           0           0           0           0

 Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
  3/49     6.73G   0.04186    0.0325  0.002924   0.07728        10       608: 100%|████████████████████████████████████████████████████████████████████████████████| 1359/1359 [2:15:52<00:00,  6.00s/it]
           Class      Images     Targets           P           R      mAP@.5  mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████| 340/340 [25:39<00:00,  4.53s/it]
             all    5.43e+03    2.41e+04           0           0           0           0

 Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
  4/49     6.73G   0.04029   0.03193  0.002385   0.07461        24       608: 100%|████████████████████████████████████████████████████████████████████████████████| 1359/1359 [2:14:47<00:00,  5.95s/it]
           Class      Images     Targets           P           R      mAP@.5  mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████| 340/340 [25:49<00:00,  4.56s/it]
             all    5.43e+03    2.41e+04           0           0           0           0

 Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
  5/49     6.74G   0.03929   0.03121  0.002087   0.07259        92       608:  78%|██████████████████████████████████████████████████████████████▋                 | 1065/1359 [1:45:58<09:14,  1.88s/it]
  5/49     6.74G   0.03914   0.03114  0.002126    0.0724        31       608: 100%|████████████████████████████████████████████████████████████████████████████████| 1359/1359 [2:13:43<00:00,  5.90s/it]
           Class      Images     Targets           P           R      mAP@.5  mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████| 340/340 [25:41<00:00,  4.53s/it]

[glenn-jocher]

@ou525 train 300 epochs and then post your results.png.

[ou525]

@glenn-jocher If training based on pre-training weights still needs so many epochs, if it can not converge quickly, then the role of pre-training weights is not lost?

[glenn-jocher]

@ou525 if you want to see the effect, then train with and without them, both to 300, and post your results.

[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.