Pytorch C++ Front

Sometimes we need an efficiency, low latency enviroment to run our model.
Also, most of the embedded system is written in C/C++, the C++ front can makeC++ developers directly use C++ to build their pytorch model as well.

You could reference the official installing guide to build the example.

The design of the C++ front-end follow the following rules:

Closely model the Python frontend in its design:
It means that most of the operations are similar to the python-front style.

For example:

Define Module and layer in python:

Class SimpleNet(nn.Module):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.linear = torch.nn.Linear(N, M)
       
    def forward(self, x):
        return self.linear(x)

Corresponding C++ implementation

struct SimpleNet : torch::nn::Module {
    SimpleNet(int64_t N, int64_t M)
        : linear(register_module("linear", torch::nn::Linear(N, M))) {}

    torch::Tensor forward(torch::Tensor input) {
        return linear(input) + another_bias;
    }

    torch::Tensor forward(torch::Tensor x){
        return linear(x);
    }

    torch::nn::Linear linear;        
};

Flexibility and User-friendliness over micro-optimization:

Usually, C++ can get more optimized efficacy but complexities in creating the program. So, PyTorch trying to let the c++ developer get more user-friendly usage.

For example, C++ can get the parameter just similar to python :
```
## python
net = SimpleNet(4, 5)
print(list(net.parameters()))

// C ++
SimpleNet net(4, 5);
for (const auto& p : net.parameters()) {
    std::cout << p << std::endl;
}
```

For detail, Official cpp_frontend had explain well.

Error when building torch c++ front

find_package error
```
The package name passed to `find_package_handle_standard_args` (torch) 
does not match the name of the calling package (Torch).
```
This should be the problem of the PyTorch c++ package. Someone said that could evolve with PyTorch 1.4 and 1.5. So you may downgrade to cmake <= 3.16 and the warning should disappear.

Reference Link
cannot find -lCUDA_cublas_LIBRARY-NOTFOUND

I did not know why CMake can not find cublas even if I set the LD_LIBRARY_PATH.
```
LD_LIBRARY_PATH=/usr/local/cuda-10.1/lib64:$LIBRARY_PATH
```
Solution: mutually create the link to user library.
```
sudo ln -s /usr/local/cuda-10.1/lib64/libcublas.so.10 /usr/lib/libcublas.so
```

Pytorch C++ Front

Error when building torch c++ front

Chia-Tse Chang