Binbin Xu(徐彬彬)

[这里](https://www.microsoft.com/en-us/research/wp-content/uploads/2013/01/Face-Alignment-by-Explicit-Shape-Regression.pdf)2.1节讲了原因，face-alignment-at-3000fps是在这篇的基础上做的，框架一模一样

没有改任何代码，就是用README中提供的数据从头训练（而不是finetune），21个epoch结束后显示准确率只有10%，实际应该更低，预测出来全部是“的的的的的”，不知道哪里有问题

我的loss好像没怎么变，从最开始60左右，到21个epoch结束的时候还是60左右，中间有没有下降不清楚，忘记重定向log了

我用的是anaconda python3.6.3, cuda9.0, cudnn 7.0, pytorch 1.0.1.post2~ 是否跟pytorch的版本有问题? @chenjun2hao 你的模型是在什么环境下得到的？我可以重新跑一下

https://github.com/dmlc/cxxnet/blob/master/doc/multigpu.md in the above documentation it says it needs to be done in the future, so we must wait

yes, inside the warp-ctc these is softmax, for example in `cpu_ctc.h` you can see [cpu softmax](https://github.com/baidu-research/warp-ctc/blob/master/include/detail/cpu_ctc.h#L79)

happend to see the same problem when binding with tf 1.1.0, though `python setup.py test` can not run correctly, but I can still use warp-ctc by installing `python setup.py install`,...

@jungomi hi, can you give your accuracy? is your implementation has similar performance as the paper?

same problem, do not know how to fix

> same problem, do not know how to fix I solved it by upgrading python 3.7 to python 3.9