search

ultralytics / yolov3

YOLOv3 in PyTorch > ONNX > CoreML > TFLite
https://docs.ultralytics.com
GNU Affero General Public License v3.0
10.18k stars 3.44k forks source link

Am I using CIOU loss wrong? #1082

Closed tinothy22 closed 4 years ago

tinothy22 commented 4 years ago

i use GIOU loss for training my custom dataset with yolov3-tiny, get 87% MAP,and below is result results then i use CIOU loss ,and get 86.8%,the results did not improve,here are the results of CIOU: results Probably CIOU is not effective? I just I set CIOU = True in giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, CIoU=True) can you give me some guidence?

glenn-jocher commented 4 years ago

@tinothy22 sounds about right. They're very similar losses.

tinothy22 commented 4 years ago

Ask a stupid question, how can I save the entire model, not just the parameters of the model. Is it to change the dictionary chkpt?

glenn-jocher commented 4 years ago

torch.save(model)

tinothy22 commented 4 years ago

Can you elaborate? I want to generate a .pth model file for visualization with netron. Where can I insert the code torch.save, can you help me?

glenn-jocher commented 4 years ago

@tinothy22 netron is pretty bad with pytorch files. You'll get better results using it directly to view the *.cfg files.

tinothy22 commented 4 years ago

Thank you very much for your answers

glenn-jocher commented 11 months ago

@tinothy22 you're welcome, but credit really goes to the YOLO community and the Ultralytics team. They've done an amazing job with the YOLOv3 ecosystem.