I just started using wandb and really liked it so far. I have a question about resetting global_step. I went through the asked questions but didn’t find something useful.
In my code, I train an RL agent and plot loss and mean reward values during training. Once the training is over, I run the agent over the environment for a number of test episodes. I need to plot the mean reward for test episodes as well. However, since there’s this global step, in the plots, the mean reward for test runs doesn’t start at zero which is annoying. How can I address this issue?
Thank you for writing in! Could you please point me to your wandb workspace where you are experinencing this behavior?
Here is a simple report containing the issue: Report for the WandB issue
Thank you for sending it over! Does it start at 100k for you instead of 0?
Yes, exactly that’s the problem I have.
Thank you for elaborating. When you are logging those metrics, do you start from the step 100k, or is there some data that you have before the 100k that just doesn’t get displayed?
initially, I log the training stuff for 100k and it’s perfect. Once the training is done, I run the policy for evaluation and want to log evaluation stuff, but since the global step cannot be reset, it starts from 100k which is annoying.
Thank you for elaborating. When you start logging your data at 100k, unfortunately there isn’t a way to get the 0-99,999 for you. I think the only potential solution here would be logging your data into a table instead of the regular wandb.log, and then graphing that table using custom charts.
Otherwise, unfortunately, it is not possible to change your global_step manually in the UI.
Hi @modanesh! Wanted to follow up with you regarding this concern!
Thanks for the info. I understand the limitation. If you could support this type of logging it would have been very helpful.
I can submit a feature request for you if you’d like
That would be great. Thanks.
No Problem! Your feature request has been submitted.