Sakya Demon

我也很纳闷为什么同样的代码会有这么多的差异。你可以试试改改 batch size 或者初始化参数的方法。

越界的时候，代码的做法是使用了 cosine间距 去代替 角间距。理论上是可行的。估计越界的情况不会很多，应该也不会影响效果。你说使用 clip ，基本上是把这个样本对给放弃了。同上，应该也不会有什么影响。不过我没有试过就是了啦。

More detail ?

Thank you! @Jack-CV

I have fixed the bug (coding mistake) but there is still one image failed. See [this](https://github.com/siriusdemon/pytorch-PCN/blob/master/result/ret_20.jpg) I think the Pytorch version is ready to use. Thank you! @Jack-CV

I haven't train on new data myself too. I just convert the model from the original repo. For training details, you can refer to the paper and implement by yourself....

Sorry, I haven't train it. see this [issue](https://github.com/siriusdemon/pytorch-PCN/issues/2)

先安装好依赖 ```sh pip3 install numpy opencv-python pip3 install https://download.pytorch.org/whl/cpu/torch-1.0.1.post2-cp36-cp36m-linux_x86_64.whl pip3 install torchvision ``` 再克隆仓库： ```sh git clone https://github.com/siriusdemon/pytorch-PCN.git ``` 完成之后进入文件夹里 ```sh cd pytorch-PCN ``` 然后执行代码 ```sh python demo.py imgs/1.jpg ```...

For a pytorch model, you can set model and input to CUDA model. e.g. ``` model.cuda() input.cuda() ``` Then it will run on GPU~

Some operations like `NMS` is time-consuming. There are lots of `for loop` in the code. Analyse the bottleneck by `profile` and rewriting some code by `Cython` may help.