search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.71k stars 7.96k forks source link

Training loss not dropping (Gaussian and GI #4252

Open peterdn1 opened 4 years ago

peterdn1 commented 4 years ago

I am trying out both Gaussian_yolo and GIoU iou_normalizer=-0.5 iou_loss=giou

Pulled and compiled repo before training Custom dataset 17 classes (using classes + 8 + 1) * number masks Used the darknet53.conv.74 weights

Started learning rate at 0.001 and avg loss went to -nan after about 1300 iterations and remained there.

Restarted: Set the learning rate to 0.0001 and avg loss is not dropping (25000 iterations sitting at ~400-440

Training with yolov3 same dataset trained well.

Can I use GIoU and Gaussian together? If so any suggestions?

Thanks

AlexeyAB commented 4 years ago

Can I use GIOU and Gaussian together?

GIoU isn't implemented in Gaussion-yolo yet.

Set the learning rate to 0.0001 and avg loss is not dropping (25000 iterations sitting at ~400-440

This is the normal loss for Gaussion-yolo networks. Look at the mAP.

peterdn1 commented 4 years ago

Ran the partially trained network with the avg loss of ~400 and results looked good, thank you.

Can the loss calculation be adjusted for graphing function makes sense for Gaussian

If you combine GIoU and Gaussian do you expect a linear improvement in mAP? (Gaussian mAP gain+GIoU gain)