How to save scaling factors and offsets in ONNX models and access them through c++ onnxruntime?

[affafj]

I want to find out how to store other user-defined parameters in a onnx model and then access them with c++ onnx runtime. Here is a detailed description.

I have the following neural network model in pytorch. It's a simple MLP with 2 hidden layers, 50 + 50 inputs and 3 outputs (because $p=q=1$):

class NARXnet1(nn.Module):
    scale = [250.0,250.0,250.0]
    offset = [0.0,0.0,0.0]
    def __init__(self,p,q,m,n):
        super(NARXnet1,self).__init__()
        self.input_fc = nn.Linear(m+n,80) # fully-connected input thrust layer
        self.actvn1 = nn.ReLU() # activation layer 1
        self.fc2 = nn.Linear(80,80) # fully-connected hidden layer
        self.actvn2 = nn.ReLU() # activation layer 2
        self.fc3 = nn.Linear(80,p+q+1) # output layer

    def forward(self,T_in,u_in): # function for forward propagation
        combined = torch.cat((T_in, u_in), 1)
        combined = self.input_fc(combined)
        combined = self.actvn1(combined)
        combined = self.fc2(combined)
        combined = self.actvn2(combined)
        outputs = self.fc3(combined)
        return outputs

Notice the attributes scale and offset in the above class? I want to use these attributes to scale and offset the output of my neural network during inference in c++. How can I do that?