Check which is the relationship of iterations and epochs
Closed this issue · 2 comments
Check which is the meaning of iterations and epochs in ray tune and wether the step() function should include explicitly the amount of epochs to run.
link to ray discussion forum post --> https://discuss.ray.io/t/relationship-of-epochs-and-training-itertions/14433
We have discussed that the function step() in the Trainable class should have an intermediate running time - not too long, not too short. This would depend on the user dataset.
Therefore, we now will have 2 variables (not tuned) in the config file.
One is step_size and it defines how many iteration of the full datasets should be done in each step.
One is epochs which is the number of times the step() function will be called.
The final number of epochs, as we intend them traditionally as one full iteration through the dataset, in training mode it will be epochs*step_size.