去看pytorch.org 的tutorial!!!
I'd say that the official tutorials are a great start (Welcome to PyTorch Tutorials). There you have a lot of examples of all the things you'll probably run into when trying to design an architecture and train it: dataloaders, NN modules, classes, forward and backward passes, optimizers, schedulers, etc.
My suggestion would be to first read all of the tutorials (and really understand them code-wise), and then implement your own architecture for a problem of your choosing. As always, MNIST is by far a really easy example which you can work with.
When you feel comfortable with the framework, I say you should visit GitHub repos regarding your topics of interest. A LOT of code from famous papers is public, and I think they have cool things from which we can learn.
As a sidenote, when having problems (technical ones or with your net) remember to visit pytorch's forums (PyTorch Forums), since there you'll probably find the answer to any problem you may be having.