Getting Started¶
SMAC needs four core components (configuration space, target function, scenario and a facade) to run an optimization process, all of which are explained on this page.
They interact in the following way:
Configuration Space¶
The configuration space defines the search space of the hyperparameters and, therefore, the tunable parameters’ legal ranges and default values.
from ConfigSpace import ConfigSpace
cs = ConfigurationSpace({
"myfloat": (0.1, 1.5), # Uniform Float
"myint": (2, 10), # Uniform Integer
"species": ["mouse", "cat", "dog"], # Categorical
})
Please see the documentation of ConfigSpace for more details.
Target Function¶
The target function takes a configuration from the configuration space and returns a performance value. For example, you could use a Neural Network to predict on your data and get some validation performance. If, for instance, you would tune the learning rate of the Network’s optimizer, every learning rate will change the final validation performance of the network. This is the target function. SMAC tries to find the best performing learning rate by trying different values and evaluating the target function - in an efficient way.
def train(self, config: Configuration, seed: int) -> float:
model = MultiLayerPerceptron(learning_rate=config["learning_rate"])
model.fit(...)
accuracy = model.validate(...)
return 1 - accuracy # SMAC always minimizes (the smaller the better)
Warning
SMAC always minimizes the value returned from the target function.
Note
In general, the arguments of the target function depend on the intensifier. However,
in all cases, the first argument must be the configuration (arbitrary argument name is possible here) and a seed.
If you specified instances in the scenario, SMAC requires instance
as argument additionally. If you use
SuccessiveHalving
or Hyperband
as intensifier but you did not specify instances, SMAC passes budget as
argument to the target function. But don’t worry: SMAC will tell you if something is missing or if something is not
used.
Warning
SMAC passes either instance or budget to the target function but never both.
Scenario¶
The Scenario is used to provide environment variables. For example, if you want to limit the optimization process by a time limit or want to specify where to save the results.
from smac import Scenario
scenario = Scenario(
configspace=cs,
output_directory=Path("your_output_directory")
walltime_limit=120, # Limit to two minutes
n_trials=500, # Evaluated max 500 trials
n_workers=8, # Use eight workers
...
)
Facade¶
A facade is the entry point to SMAC, which constructs a default optimization pipeline for you. SMAC offers various facades, which satisfy many common use cases and are crucial to achieving peak performance. The idea behind the facades is to provide a simple interface to all of SMAC’s components, which is easy to use and understand and without the need of deep diving into the material. However, experts are invited to change the components to their specific hyperparameter optimization needs. The following table (horizontally scrollable) shows you what is supported and reveals the default components:
Black-Box |
Hyperparameter Optimization |
Multi-Fidelity |
Algorithm Configuration |
Random |
Hyperband |
||
---|---|---|---|---|---|---|---|
#Parameters |
low |
low/medium/high |
low/medium/high |
low/medium/high |
low/medium/high |
low/medium/high |
|
Supports Instances |
❌ |
✅ |
✅ |
✅ |
❌ |
✅ |
|
Supports Multi-Fidelity |
❌ |
❌ |
✅ |
✅ |
❌ |
✅ |
|
Initial Design |
|||||||
Surrogate Model |
Not used |
Not used |
|||||
Acquisition Function |
Not used |
Not used |
|||||
Acquisition Maximizer |
|||||||
Intensifier |
|||||||
Runhistory Encoder |
|||||||
Random Design Probability |
8.5% |
20% |
20% |
50% |
Not used |
Not used |
Note
The multi-fidelity facade is the closest implementation to BOHB.
Note
We want to emphasize that SMAC is a highly modular optimization framework. The facade accepts many arguments to specify components of the pipeline. Please also note, that in contrast to previous versions, instantiated objects are passed instead of kwargs.
The facades can be imported directly from the smac
module.
from smac import BlackBoxFacade as BBFacade
from smac import HyperparameterOptimizationFacade as HPOFacade
from smac import MultiFidelityFacade as MFFacade
from smac import AlgorithmConfigurationFacade as ACFacade
from smac import RandomFacade as RFacade
from smac import HyperbandFacade as HBFacade
smac = HPOFacade(scenario=scenario, target_function=train)
smac = MFFacade(scenario=scenario, target_function=train)
smac = ACFacade(scenario=scenario, target_function=train)
smac = RFacade(scenario=scenario, target_function=train)
smac = HBFacade(scenario=scenario, target_function=train)