nn.Parameter inside class, but the module doesn't seem to contain the nn.Parameters

1

I’m creating a class, details are not so important just the fact that in init there are conv layers and nn.Parameters:

class MyClass(nn.Module):
    def __init__(self, args):
        super(MyClass, self).__init__()
        nc = 100
        self.teta0 = nn.Parameter(torch.cuda.FloatTensor(1,nc),requires_grad=True)
        self.teta1 = nn.Parameter(torch.cuda.FloatTensor(1,nc),requires_grad=True)
        self.teta2 = nn.Parameter(torch.cuda.FloatTensor(1,nc),requires_grad=True)
        self.conv  = nn.Conv2d(1,3,3)
    def forward(self,a,b): 
        first  = self.teta2*(a-self.teta1)
         ....
        return loss

Now I’m instantiating the class and print

        self.model= MyClass(args)
        print(self.model)

I’m getting:

MyClass (
(conv): Conv2d(1, 3, kernel_size=(3, 3), stride=(1, 1))
)

It’s seems like the class is not consisting of those nn.Parameters.
Just to makes things clear, self.conv is there because I had this problem and I wanted to see if the conv layer can be seen.
Any idea what’s going on, where I’m mistaking and what can lead to such an error?
Thanks!

2

Conv2d is a module. In the source for nn.Conv2d, Conv2d inherits from _ConvNd which in turn inherits from Module. On the other hand, nn.Parameter inherits from torch.Tensor. From what I understand, this implies that the parameters, although learnable, are not part of the Model to begin with and any operation involving it is dynamically computed. This is why initializing torch.ones in the init section does not add that variable to the model.