Nevermore comments

Results 25 comments of


                                            Nevermore

when distilling Faster RCNN with FPN, do we need to firstly assign gt box to different FPN level and then generate imitation region? Or for every FPN level, we generate imitation region according to all gt boxes?

> I met with some problems when add your proposed imitation loss to faster rcnn with fpn, I assign gt in this way: one a certain level, for a gt...

请问有没有试过backbone 为resnet的蒸馏

I try to use res50 based FPN to teach res18 based FPN, but the mAP is not so good. Is that normal?

A higher result than paper, but i don't know why.

Hi, Is there any new progress on this result? The ViP_s7 got the 85.16 ImageNet top-1 or Real top-1?

训练报错

如果要用KD的话，应该是 ```python tools/train_kd.py```

ImportError: cannot import name 'Runner_kd' from 'mmcv.runner'

感谢@ruiningTang 的回复，因为我在mmcv/runner 下面加了个runner_kd.py 所以直接pip install mmcv是没有我自己加的那个.py的所以会import error。是我没写清楚“从source编译的含义”，就是clone下来之后把runner_kd.py放进去一起编译就行了。

ImportError: cannot import name 'Runner_kd' from 'mmcv.runner'

用```python -m torch.distributed.launch --nproc_per_node=8 tools/train.py```这个代码训练也会报这个错吗

WORKDIR = './ work_dirs/hitdet_1x/ '

这个是mmdetection留下的接口，我没删掉，没其他的目的。

请问可以提供faster-rcnn-v2-374.pth吗？

这个可以根据[MMDetection](https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn)的直接训出来，很稳定基本是37.4，也可以用train.py直接训一个faster-rcnn，我的.pth太久远了不知道还能不能找得到。

About the train log of on coco dataset

Sorry, I can't find the training logs now. I think the loss weight (λ) should be 4 and 16 for the foreground and background, and I use KD for backbone...

配置问题

这个问题是依赖包mmcv没安装好，可以参考[MMDet官方github](https://github.com/open-mmlab/mmdetection/blob/master/docs/get_started.md)。