Is it necessary to save loss while saving checkpoint?
Is it an important parameter?
If I save checkpoint with model.state_dict(), optimizer.state_dict() and scheduler.state_dict(), when I restart training from checkpoint, I see first epoch has different loss value
loss isn’t a parameter but it is computed, so you don’t need to save it except for informational purposes.
It would be expected that the loss is different after loading and running the model again, just like it would be if you had run the original model right after saving it.
Thanks @Tom for the detailed answer
What are the important parameters do we need to save in checkpoint?