I have a few questions regarding the hyperparameter sweeps from Python.
I am wanting to essentially start a few tmux sessions on my server, and connect them all to the same sweep agent, but no keyword in the sweep_config (that i have found) allow me to connect to a specific sweep ID, and rather just a sweep name that doesnt connect to the same sweep, but just makes multiple sweeps of the same name. If this possible or strongly advised against due to computational usage or similar?
Furthermore, sweeps take up a great deal of storage requirements due to saving all the models, is it possible to store the model file from the best model only, while keeping the statistics from all the models for plots and interpretation? This would allow me to keep the great information gathered from sweeps, while not taking up 100+ GB from a single sweep.
Hi @tjobbertjob , for your first question. Try creating a new agent and passing your existing (running) sweep id to the new agent.
As for your questions regarding the model saving, you would need to either write some custom saving behaviour to check if model is the best model in the sweep thus far (by using the public api) or you could write some script that asynchronously and programmatically deletes all but the top-k models every every so often.
When and where should i add the sweep id? in the sweep_id config dictionary, or in the wandb.agent? and what is the key for the id?
I found the issue, i was trying to create a new wandb.sweep(config, project, entity) and pass the ID into the config dictionary, but instead i just needed to take the ID directly, and just do sweep_id = sweep_id_string which worked.
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.