帅约科维奇
帅约科维奇
What is your mAP? I also have tried to train it but result in a very low accuracy.
We don't manually restrict each level feature to detect objects within a certain range of sizes. The selection of which FPN level each GT should be assigned to is decided...
Thanks for your interests in our work. We will try to apply dw on yolov6 in the coming days.
Which version of PyTorch and mmdetection do you use?
This error may arise from the FPN code of MMDetection. Similar problems can also be found at https://github.com/open-mmlab/mmdetection/issues/7933. My environment is: pytorch: 1.9.0+cuda11.1, mmdet: 2.23.0, mmcv:1.4.8.
> > Have you ever trained on coco for these models? I try to train coco with original SSD. However, the loss keeps increasing especially for the localization loss from...
recall = true_positive / total_gt, so the neg prediction will not affect the recall.
In the code, the classificaiton loss consists of cls_loss obatained by p_pos_weight and cls_neg_loss obtained by p_neg_weigth as shown [here](https://github.com/strongwolf/DW/blob/27f50e5bf8d9ea021d5fe1a7a7e47e06bc5db208/dw_head.py#L181) and [here](https://github.com/strongwolf/DW/blob/27f50e5bf8d9ea021d5fe1a7a7e47e06bc5db208/dw_head.py#L175).
In the code, t = p_cls * p_loc. p_loc = torch.exp(-5*reg_loss) = torch.exp(-5* (-torch.log(iou))) = iou**5 if we use IoU loss. So t = p_cls * iou**5 which is the...
I used 1.4.8 for mmcv and 2.28.0 for mmdet.