MLOS
MLOS copied to clipboard
microsoft
MLOS is a project to enable autotuning for systems.
Often times we will run with `teardown: false` and tweak experiments for different variations. This means that after a while we may have stale resources running and eating resources in...
And maybe even automatically include the client IP if possible.
Sometimes after reconfiguring a machine it ends up in a strange bad state (e.g., OOM and swapping so all commands are slow). We should check for this and issue a...
```[tasklist] ### Tasks - [ ] Create resource group to hold resources - [ ] Create service account to run things - [ ] Stash service account's auth token in...
Some system parameters require reloading data or restarting to take effect. We should make sure that we have examples at least or new abstractions for those.
When errors are encountered during config loading and setup it can sometimes be hard to identify which config they came from. For instance, often time we just have a simple...
Right now we rely on covariant tunable groups to associate tunables with the environment. In real experiments, we oftentimes want to use only a selection of parameters from a covariant...
Options: - max_cost - max_time - max_iterations (currently available) - "converaged" - e.g., max_change over M < max_iterations steps
Metadata introduced in #411 should be passed to mlos_core optimizers as a meta field of ConfigSpace hyperparameters. We also need unit tests to make sure the conversion is correct
See Also: - 38560231e6b26a81642b11d22ed73dd8e89a2f4a from #392 - https://github.com/automl/ConfigSpace/commit/e1212277932217fc3898be17255a7d4a71904b30