pytorch多进程运行模型，报错：报错 RuntimeError: Cowardly refusing to serialize non-leaf tensor which requires_grad, since autograd does not support crossing process boundaries.

具体报错信息：

报错 RuntimeError: Cowardly refusing to serialize non-leaf tensor which requires_grad, since autograd does not support crossing process boundaries. If you just want to transfer the data, call detach() on the tensor before serializing (e.g., putting it on the queue).
[W412 09:35:34.731165326 CudaIPCTypes.cpp:16] Producer process has been terminated before all shared CUDA tensors released. See Note [Sharing CUDA tensors]

外网上关于这个报错信息的说法也是各种各样，国内的呢，也是基本见不到这方面的资料，于是就自己研究，发现了解决方法。

这个问题的出现其实就是在python的多进程中开启pytorch模型，如果是在子进程的__init__函数中创建模型，那么python就会把这个模型进行序列化，然后就会报错，解决方法就是把子进程中的神经网络初始化拿出来放到非__init__函数中就可以了。

具体：

改成：