JimLee1996

Half precision is also helpful to reduce memory consumptiuon. Try use `tensor.half()` and `model.half()`.

3DCNN is supported since I've already tested. Please paste your error message and code here.

路由器硬件太差了，跟不上

Sure! Since I'm not a specialist in swift, it may be tough for me to further develop this project. However, we can work it out together and make it swifter...

这个和我最初做的menubar图标好像啊，哈哈。当时觉得字太小眼神不好就没用。我的建议是可以让社区投票征集一下，然后我们再进行变动。

Try to delete launch agents of trojan? 1. stop and quit trojanx 2. delete launch files of trojan ``` Last login: Mon Mar 2 11:53:13 on ttys002 ➜ ~ cd...

找到trojan的pid，把trojan的二进制进程kill掉。或者重启之后就好了

Maybe the config is to blame. I've hard-coded some parameters in the config. Please check these [lines](https://github.com/JimLee1996/TrojanX/blob/a1985a0de5502e8f5ed4f623cd9b9d61d7c9e8e7/TrojanX/LaunchAgentUtils.swift#L114-L117) and paste the corresponding config of your server.

作者居然把输入的shape都能搞错，所以可以说与训练的模型是废的，没法测试的。 ```python input_tensor = Input(name='the_input', shape=(96, 64, 3)) ``` 应当为 ```python input_tensor = Input(name='the_input', shape=(64, 96, 3)) ``` resize的时候是传入图像的(W, H) 当成tensor看的时候是(H, W) 要么就不开源，开源就把code base做好，不要随便弄弄就跑路。

服务器尽量保持在线就好~但我有个疑问，就是建立了P2P后，流量是怎么走的，有大神解释一下么