Difference between yolov7.pt and yolo7_training.pt

[DP1701]

Hi all, what is the difference between yolov7.pt and yolo7_training.pt?

[holger-prause]

• For transfer learning you would use yolo7_training.pt
• For training from scratch you would use random weights - i.e train with - weights ''
• For doing prediction on a pretrained model you would use yolov7.pt

What the actual difference between yolo7_training.pt and yolo7.pt is - i actually don't really know. I could just speculate that yolov7.pt contains some things you don't need for training - like a nms layer on the end and reparametrized layers - means some of its trainable layers were converted into convolutional layers to speed up inference time.

However - training and using both for transfer learning seems somewhat fine. I would advise you to use the yolo7_training.pt for transfer learning.

If anyone with more knowledge could comment on this - this would be great.

[akashAD98]

@holger-prause @DP1701 @WongKinYiu
yolov7_training.pt you can use for transfer learning &
yolov7.pth is parameterized weight of yolov7_training.pth(used for deployment purposes)
lots of people have doubts regarding reparameterised so I submitted pr #552

[DP1701]

@akashAD98 Transfer learning also requires the —freeze command, or am I wrong?

[koshinryuu]

very strange, when I train and test a very small dataset coco128

• yolov7_training.pt after 100 the precision and recall is very bad, about 0.002 each

  • python train.py --img 640 640 --cfg cfg/training/yolov7.yaml --weights 'yolov7_training.pt' --hyp data/hyp.scratch.p5.yaml

• yolov7.pt after 30 epochs is very well , about 0.8map

  • python train.py --img 640 640 --cfg cfg/training/yolov7.yaml --weights 'yolov7.pt' --hyp data/hyp.scratch.p5.yaml

1. why did the result with yolov7_training.pt is so bad ??
2. could I use yolov7.pt for training a custom dataset?