Feedback about Quickstart

[rjdamiano]

Page: https://docs.pytorch.org/tutorials/beginner/basics/quickstart_tutorial.html

The example training loop under the Optimizing the Model Parameters section is not using the standard order of zero_grad(), backward(), step().

As written currently in the documentation:

# Backpropagation
loss.backward()
optimizer.step()
optimizer.zero_grad()

The same issue exists in the Optimizing Model Parameters: page: https://docs.pytorch.org/tutorials/beginner/basics/optimization_tutorial.html

cc @albanD @jbschlosser

[albanD]

I don't think it really matters for simple training loop like this.
And if you're doing microbatching or more advanced techniques, neither will work anyways.
So I think it's ok to keep as is

[rjdamiano]

My argument for the change is that in the context of the documentation section - "Getting Started with PyTorch / Learn the Basics" - it could be confusing for a beginner, especially when a later documentation section explains them in the standard order: https://docs.pytorch.org/tutorials/beginner/basics/optimization_tutorial.html#optimizer