How could I get the result out and store it

I use to do it like this: net(im).detach().cpu().numpy() until I got to know detach() allows two tensors have a same data which I believe means that if I change one another will be modified.

What is the correct way to store the result of a model, for example, the loss or the embeds ?

Yes, detach() uses the same underlying data, but detaches the variable from the computation graph.
To store tensors, you could simply use, file).
If you want to save your model parameters, I would recommend to store the state_dict.
Have a look at the Serialization Semantics for more information.