There are two ways to use SMAC, either over the commandline or within your Python-code.
Either way, you need to provide the target algorithm you want to optimize and the configuration space, which specifies the legal ranges and default values of the tunable parameters. In addition, you can configure the optimization process with the scenario-options.
The most important scenario-options are:
- run_obj: Either quality or runtime, depending on what you want to optimize.
- cutoff_time: This is the maximum time the target algorithm is allowed to run.
- wallclock_limit, runcount_limit and tuner-timeout are used to control maximum wallclock-time, number of algorithm calls and cpu-time used for optimization respectively.
- instance_file, test_instance_file and feature_file specify the paths to the instances and features (see file-formats)
For a full list, see scenario-options.
SMAC is called via the commandline with the following arguments:
python smac --scenario SCENARIO --seed INT --verbose_level LEVEL --mode MODE
- scenario: Path to the file that specifies the scenario for this SMAC-run.
- seed: The integer that the random-generator will be based upon. Default: 12345
- verbose_level: in [INFO, DEBUG], specifies the logging-verbosity. Default: INFO
- mode: in [SMAC, ROAR]. SMAC will use the bayeasian optimization with an intensification process, whereas ROAR stands for Random Online Adaptive Racing. Default: SMAC
- restore_state: A string specifying the folder of the SMAC-run to be continued. Assuming exactly the same scenario, except for budget-options.
In the scenario file, there are two mandatory parameters: The algo-parameter
defines how SMAC will call the target algorithm. Parameters will be appended to the call
-PARAMETER VALUE, so make sure your algorithm will accept the parameters in this
form. Read more in the section on target algorithms.
The paramfile-parameter defines the path to the PCS-file,
which describes the ranges and default values of the tunable parameters.
Both will interpret paths from the execution-directory.
Currently, running SMAC via the commandline will register the algorithm with a Target Algorithm Evaluator (TAE), that requires the target algorithm to print the results to the console in the following format (see Branin):
Result for SMAC: <STATUS>, <runtime>, <runlength>, <quality>, <seed>, <instance-specifics>
If a SMAC-run was interrupted or you want to extend its computation- or
time-limits, it can be restored and continued.
To restore or continue a previous SMAC-run, use the
--restore_state FOLDER-option in the commandline. If you want to increase
computation- or time-limits, change the scenario-file specified with the
--scenario SCENARIOFILE-option (not the one in the folder to be restored).
Restarting a SMAC-run that quit due to budget-exhaustion will do nothing,
because the budget is still exhausted.
Changing any other options than *output_dir*, *wallclock_limit*, *runcount_limit* or
*tuner-timeout* in the scenario-file is NOT intended and will likely lead
to unexpected behaviour!
For an example of restoring states from within your Python code, there is an implementation with the Branin-example in “examples/branin/restore_state.py”.
Usage in Python¶
The usage of SMAC from your Python-code is described in the SVM-example. Scenario and configuration space are both build within the code. The target algorithm needs to be registered with a Target Algorithm Evaluator (TAE), which communicates between SMAC and the target algorithm. To optimize a function, you can instantiate ExecuteTAFuncDict or ExecuteTAFuncArray. In that case, the algorithm needs to return a cost, representing the quality of the solution, while time- and memory-limits are measured and enforced by Pynisher, so no wrapper is needed for your algorithm here.
- ExecuteTAFuncDict: The target algorithm is called with a dict-like configuration and optionally with seed and instance, returning either the loss as a float or a tuple (loss, additional information).
- ExecuteTAFuncArray: The target algorithm is called with an array-like configuration and optionally with seed and instance, returning either the loss as a float or a tuple (loss, additional information).