Yipeng Gao

You can try build the model in the ModelEmaV2 class directly, that works for me. Eg: forward the config into the Ema class and build the model in the "__init__"...

Hi, can you share your GPU type and errors? I used Python 3.7 and pytorch 1.7. Such an env may be relatively old for now, but I think the codebase...

Hi 感谢对这个文章的关注！1) 这里的layer代表利用yolov5 backbone的哪一层输出特征来进行MMD alignment； 2）感谢建议！我会在一周左右在readme中做更详细的说明和论文涉及的部分解释。

hello heyuhhh and zhouzheng1123 非常不好意思，train.py可以先用这个https://github.com/Hlings/AsyFOD/blob/main/train_files_for_abl/train_select_ins_dissim_to_sg-sim-tar_layer9.py 能够reproduce结果。batchsize在Cityscapes上是12 or 8，在Sim10K到C和KITTI到C都是32 or 16。Batchsize在Foggy Cityscapes上比较敏感，在其他数据集上性能差的不会很多（根据我之前的实验经验）。 如果还有问题可以通过邮箱[email protected]联系我，我应该能够更快回复。

> 感谢作者耐心的回复~ 我现在在复现sim10k->city的实验结果，但是使用上面说的train_select_ins_dissim_to_sg-sim-tar_layer9.py只能得到58%mAP左右的结果。 我使用的数据集是从百度网盘下载下来的，batchsize是24，数据增强是mm4.yaml，其他的代码基本上没有动过。请问一下作者大大还需要什么需要注意的地方嘛？ 该问题已经和提问者沟通解决。因为小样本，数据采样会带来一定的性能波动，多次采样后的性能和论文里报告的性能基本一致，对应采样的数据和在sim10k->city的训练设置已在readme中进行更新。

HI. Sorry for the late reply due to so many busy things :(. You can refer [here](https://github.com/Hlings/AsyFOD/blob/c65a90c79b4a0e55a706c5af04e0ce9d1ecb881e/utils/MMD.py#L30C5-L30C20) for the code to extract ins-level features. I find the train.py file misses...

Hi! Sorry for the late reply due to the CVPR rebuttal. Could you please provide the error with log? I think a potential modification exists in the Ranker (nn.Model) that...

Hi. Thanks for your reminder! I might have mixed up the train.py file while organizing. I have corrected the code following a similar type [here](https://github.com/Hlings/AsyFOD/blob/main/train_files_for_abl/train_select_img_dissim_to_sim-tar_layer9.py#L356). Please try again and feel...

Hi I ran the experiments on Python 3.7 or 3.8 and pytorch 1.7. If the env can run the YOLO v5, then it can also run this repo. And, btw,...

Thank you for your effort and valuable advice! I have fixed this bug.