Lue Fan

Could you be more specific and let us know which part you do not know how to deal with?

We use the visualization tools in [SimpleTrack](https://github.com/TuSimple/SimpleTrack#demo-and-api-example). Hope it helps.

I do not fully understand your question. Would you mind making it more clear? It would be fine to use Chinese. 对于第一个问题: 为了去掉被扔掉的token, 在第二次划分之前会使用keep_mask_s0过滤一遍batch_win_inds_s1. 您后面的问题我没太理解．

扔掉两次是为了降低实现的复杂度，这样之后两次attention的输入都是一样的．　30/60/100是指会用零值padding到30/60/100, 实际上有效的token数本来就是不到30/60/100的（并且这里还没有真正进行padding），所以再扔掉一点也无妨，最终在`get_flat2win_inds_v2`函数中还是会把它们padding到30/60/100的．

Yes, see the last paragraph of Usage in README. Let me know if you have further questions.

My suggestions: 1. Make sure you could train other models like PointPillars. 2. Using single GPU non-distributed training and CUDA_LAUNCH_BLOCKING=1 to locate the problem.

Sorry for late reply. According to the log, the learning rate seems too large, which may lead to fail to converge.

To be honest, in the CenterHead version, the performance of the extremely large vehicle is not so good due to the center feature missing. We addressed this problem in our...

@deepmeng Could your update it again :) ?

Thanks for using. Such `UndefinedSymbol` error is likely caused by incompatible library versions. But I am not sure what goes wrong. Here is my log which contains the related information...