Open xuhao-anhe opened 1 year ago
ChannelWiseDivergence
is a response-based kd for semantic segmentation. Please use kl_div for cls head, it should work. Besides, if you want to do distillation on Reg heads, you can try it with SmoothL1 Loss.
ChannelWiseDivergence
is a response-based kd for semantic segmentation. Please use kl_div for cls head, it should work. Besides, if you want to do distillation on Reg heads, you can try it with SmoothL1 Loss.
Thank you very much for your answer. I would like to ask if there is a config file for reference。
ChannelWiseDivergence
is a response-based kd for semantic segmentation. Please use kl_div for cls head, it should work. Besides, if you want to do distillation on Reg heads, you can try it with SmoothL1 Loss.
Hello, I'm changing the loss function of loss_bbox to SmoothL1 Loss,but the final effect is not improved. Is the modification made in loss_bbox.I look forward to your reply. Thank you very much.
How to modify your configuration file, can you teach me?
Describe the question you meet
I use the CWD method,When resnet50 is used to distill resnet18, the training accuracy of the teacher's network is 80%, but the network accuracy after distillation is only 46%. What is the reason
Post related information
pip list | grep "mmcv\|mmrazor\|^torch"
[here]2022-09-02 09:53:28,087 - mmdet - INFO - Evaluating segm... 2022-09-02 09:53:30,272 - mmdet - INFO - Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.372 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.467 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.417 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.743 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = -1.000 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = -1.000 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.746 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.746 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.746 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.746 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = -1.000 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = -1.000
mmrazor
folder. [here]