How to update model only every nth timesteps?

1

The goal is to accumulate gradients and then on the N timestep update the model. Not sure how to do it?
I’m not sure this works:
On every timestep call loss.backward() and then on N iteration: call optimizer.step(); optimizer.zero_grad()

Would this work or would the gradients calculated by loss.backward() be overwritten every time step?

2

Call loss.backward() every time you want to accumulate gradients (they will be summed up) and afterwards call step() on the optimizer you’re using to update the model’s parameters.