cizhenshi
cizhenshi
I'm sorry, I haven't come across this kind of error. It seems like there's a category number mismatch. Have you modified our configuration file?
In order to avoid the insufficient GPU memory caused by too many proposals in the allocation stage, we migrated the part of proposal allocation to the CPU. This error might...
I used 32G RAM for my training, and I'm not sure how much memory it need as least.
We changed the default dataloader and added some new parameters. Just copying the configuration file is not enough. You also need to change the corresponding dataloader.
Not the torch.utils.dataloader. Did you use a custom dataset? In the ./mmdet/datasets/coco.py or any other dataset? My configs needs to be used with ./mmdet/datasets/coco_r2.py.
If you want to use the other settings in the configs except dataset, you need to change the 171-line dataset_type = 'CocopairDataset_r2' in the config file. You can refer to...
时间问题最后没有使用randomvflip。可以看一下测试时增强这个trick
mmdetection 是通过字典来读取配置文件并进行初始化的，出现这个错误有可能是在模块初始化的时候有一些判断语句判断失败了。首先确定出错的位置是在那个文件，然后检查一下出错的地方的逻辑，修改一下对应的代码，应该就能解决问题，具体的位置我也记不清楚了。
实际上确实没有验证环节，主要因为离线验证并不准确，如果你需要离线验证的话，可以自己构建一个验证集进行验证，eval.py是有验证功能的，但是输入可能要根据你自己的数据调整一下，详细的看代码吧。 | | 肖芬瑞 | | ***@***.*** | 签名由网易邮箱大师定制 在2021年12月4日 ***@***.***> 写道： 您好，我按照您的代码，跑通后发现每一代训练之后并没有相应的测试验证环节（是我运行出错了么），如果需要，是在eval.py文件实现的吗 — You are receiving this because you are subscribed to this thread. Reply to this email directly,...
I used the default setting in mmdetection. According to my experience , frozen the first stage can achieve better results.