PyTorch Template and Tips

Template

Throughout my graduation project, company internships, and part-time work, I used both TF and PyTorch but never settled on a template. I’d add and split directories and scripts haphazardly as needed.

Templates aren’t a silver bullet, of course. Like any template, they rise and fall in popularity. But I think there’s nothing more efficient than starting development with some structure in place.

https://github.com/victoresque/pytorch-template

This example template is worth referencing not just for the template itself but also for the various implementation techniques it contains.

getattr

https://technote.kr/249

Lets you access an object’s attributes by name.

1
class Person:
2
    def __init__():
3
        self.name = 'a'
4
jack = Person()
5
getattr(jack, 'name')
6
>>> 'a'

Accessing an object’s attributes is always hardcoded. For instance, to access Jack’s name, you must write:

1
jack.name

When you want to use attributes dynamically, getattr comes in handy. It lets you modify only a config file like config.json instead of changing code every time an attribute changes.

abstract

In Java, C++, etc., you put abstract before the function name to define an abstract method. In Python, you use a decorator instead.

1
@abstractmethod
2
    def _train_epoch(self, epoch):
3
        """
4
        Training logic for an epoch
5

6
        :param epoch: Current epoch number
7
        """
8
        raise NotImplementedError

You can’t just define it without raising NotImplementedError — you need to raise the error to make the abstraction enforceable.