Wei Zhou

In my opinion, when training without SSL, you tested every snapshot on 'val' split of Cityscapes and set the best snapshot iteration as the "Early Stopping" hyper-parameter.

But if you have not got the performance of 75000 , 85000, 90000, how did you know that the best results is shown around 80000.

See [this issue](https://github.com/pytorch/extension-cpp/issues/6#issuecomment-424948254). It clarifys why this happened. Your can simply upgrade your Pytorch, the error will dispear.

同炸，很容易产生幻觉或大量重复句子

可以。具体问题请邮箱联系，git 看得比较少，基本没有怎么投入时间维护

> I setup the lib by "python setup.py install" and then run "python test_GPUpool.py". And I got an error: > > ImportError: /home/xidexia/anaconda3/lib/python3.6/site-packages/suppixpool_CUDA-0.0.0-py3.6-linux-x86_64.egg/suppixpool_CUDA.cpython-36m-x86_64-linux-gnu.so: undefined symbol: __cudaPopCallConfiguration See if .../site-packages/torch/utils/cpp_extensions.py has...

> make gcc(++) version on 5.4.0 在 RoIPooling.pytorch 和其他一些问题下也见过你啊，小哥你真是无处不在啊