-
今天阅读您的论文,提到ciou loss中 v对h和w求偏导时的问题。
由于w和h可能会非常小,所以为避免梯度爆炸您将1/(w^2+h^2) 替换成了1。
请问这在pytorch、tf这类自动求导的框架中如何实现?
我看您的pytorch实现是增加了条件S:
```
S = (iou>0.5).float()
alpha= S*v/(1-iou+v)
```
是否能理解为,使用Io…
KoapT updated
3 years ago
-
Hi Guys,
I guess there is an instability in the loss function due to a division by zero. can you please confirm that?
`def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False…
-
作者您好,我在用训练自己的数据的时候,在box_ciou()计算loss_loc的时候,出现了nan,追了一下,发现是在yolo_training.py的406-407行,在计算predict_boxes的时候,计算的w和h中,出现了inf,导致后面计算loc_loss的时候,出现了nan,导致后面的训练停止。
![QQ截图20210322135239](https://user-images.…
-
### Issue Type
Documentation Feature Request
### Source
source
### Keras Version
Keras 3
### Custom Code
Yes
### OS Platform and Distribution
_No response_
### Python ver…
2649 updated
4 months ago
-
I am trying to understand the processing of bounding boxes rn. Since yoloworld is based on yolov8 and it does not rely on pre-defined anchors, it directly predicts the bounding boxes without anchors.…
-
I train my YOLOv4 on colab.
Training and Testing is working fine, except using map to calculate mAP.
This is my command
`! cd darknet; ./darknet detector map /content/tumor/cfg/tumor.data /content/…
-
Thanks for the code,
I have the issue :
Unused field: 'iou_normalizer = 0.5'
Unused field: 'cls_normalizer = 1.0'
Unused field: 'iou_loss = ciou'
Unused field: 'nms_kind = greedynms'
Unused …
-
![Captura](https://user-images.githubusercontent.com/48415586/88299945-eece9000-ccd0-11ea-915d-cecb40abf0eb.PNG)
Hey im having this issue after training an object detector for 2 classes
`[net]
…
-
## 🚀 Feature
A loss functions API in torchvision.
## Motivation
The request is simple, we have loss functions available in torchvision
E.g. `sigmoid_focal_loss` , `l1_loss`. But these are qu…
-