Hi,
Is there any difference between changing the learning rate via param group vs creating a new object itself with a reduced learning rate as parameter?
Hi,
Is there any difference between changing the learning rate via param group vs creating a new object itself with a reduced learning rate as parameter?
Assuming “object” refers to an optimizer
here: yes, it could make a difference especially if the optimizer track internal states, as they would be kept in the former and reset in the latter case.