Badges 14 × Eureka!
Thanks ! Yeah, I finally figured it out by looking a bit more at the source code of the clearml package
Also in the source code of the Python module (
clearml/automation/optuna/optuna.py ) it looks like the behavior of randomly choosing a value between the specified limits is hardcoded. Is there a way to use a smarter/another sampler, like GridSampler for example to try every possible combination?
For example, when I use this set of hyperparameters:
DiscreteParameterRange('General/activation', values=["relu", "sigmoid"]), UniformParameterRange('General/dropout_prob', min_value=0.1, max_value=0.2, step_size=0.1), UniformIntegerParameterRange('General/n_epochs', min_value=5, max_value=15, step_size=5),
and set the
total_max_jobs to 223=12, some combinations are not tried and some other are tried multiple times
Oh, you're right, that's exactly what I was looking for. Thanks!
You probably need
--network=host , with the 2 dashes
You're welcome 🙂
It's certainly because since the string
--network=host begins with
-- , the argument parser is considering it as an option as well, and not as the value of the previous option (
--docker_args ), so you need to escape it somehow for the parser to understand that you're actually trying to pass the value for the