Add LogUnifomParameterRange for hyperparameter optimization with optuna (#462)

* Add LogUnifomParameterRange for hyperparameter optimization with optuna

* Add self.base and changed OptimizerOptuna to correctly create log uniform hyperparams.

* Add docstring to LogUniformParameterRange class.

Co-authored-by: Pereyra, Diego <9011013@ulta.com>

clearml/automation/__init__.py

@@ -1,4 +1,5 @@
-from .parameters import UniformParameterRange, DiscreteParameterRange, UniformIntegerParameterRange, ParameterSet
+from .parameters import (UniformParameterRange, DiscreteParameterRange, UniformIntegerParameterRange, ParameterSet,
+                         LogUniformParameterRange)
 from .optimization import GridSearch, RandomSearch, HyperParameterOptimizer, Objective
 from .job import ClearmlJob
 from .controller import PipelineController
@@ -6,5 +7,5 @@ from .scheduler import TaskScheduler
 from .trigger import TriggerScheduler
 
 __all__ = ["UniformParameterRange", "DiscreteParameterRange", "UniformIntegerParameterRange", "ParameterSet",
-           "GridSearch", "RandomSearch", "HyperParameterOptimizer", "Objective", "ClearmlJob", "PipelineController",
+           "LogUniformParameterRange", "GridSearch", "RandomSearch", "HyperParameterOptimizer", "Objective",
-           "TaskScheduler", "TriggerScheduler"]
+           "ClearmlJob", "PipelineController", "TaskScheduler", "TriggerScheduler"]

clearml/automation/optuna/optuna.py

@@ -2,7 +2,8 @@ from time import sleep
 from typing import Any, Optional, Sequence
 
 from ..optimization import Objective, SearchStrategy
-from ..parameters import (DiscreteParameterRange, Parameter, UniformIntegerParameterRange, UniformParameterRange)
+from ..parameters import (DiscreteParameterRange, Parameter, UniformIntegerParameterRange, UniformParameterRange,
+                          LogUniformParameterRange)
 from ...task import Task
 
 try:
@@ -193,7 +194,10 @@ class OptimizerOptuna(SearchStrategy):
         # type: () -> dict
         cs = {}
         for p in self._hyper_parameters:
-            if isinstance(p, UniformParameterRange):
+            if isinstance(p, LogUniformParameterRange):
+                hp_type = 'suggest_float'
+                hp_params = dict(low=p.base**p.min_value, high=p.base**p.max_value, log=True, step=None)
+            elif isinstance(p, UniformParameterRange):
                 if p.include_max and p.step_size:
                     hp_type = 'suggest_discrete_uniform'
                     hp_params = dict(low=p.min_value, high=p.max_value, q=p.step_size)

clearml/automation/parameters.py

@@ -171,6 +171,54 @@ class UniformParameterRange(Parameter):
         return [{self.name: v} for v in values]
 
 
+class LogUniformParameterRange(UniformParameterRange):
+    """
+    Logarithmic uniform randomly sampled hyper-parameter object.
+    """
+
+    def __init__(
+            self,
+            name,  # type: str
+            min_value,  # type: float
+            max_value,  # type: float
+            base=10,  # type: float
+            step_size=None,  # type: Optional[float]
+            include_max_value=True  # type: bool
+    ):
+        # type: (...) -> ()
+        """
+        Create a parameter to be sampled by the SearchStrategy
+
+        :param str name: The parameter name. Match the Task hyper-parameter name.
+        :param float min_value: The minimum exponent sample to use for uniform random sampling.
+        :param float max_value: The maximum exponent sample to use for uniform random sampling.
+        :param float base: The base used to raise the sampled exponent.
+        :param float step_size: If not ``None``, set step size (quantization) for value sampling.
+        :param bool include_max_value: Range includes the ``max_value``
+
+            The values are:
+
+            - ``True`` - The range includes the ``max_value`` (Default)
+            - ``False`` -  Does not include.
+
+        """
+        super().__init__(name, min_value, max_value, step_size=step_size, include_max_value=include_max_value)
+        self.base = base
+        
+    def get_value(self):
+        """
+        Return uniformly logarithmic sampled value based on object sampling definitions.
+
+        :return: {self.name: random value self.base^[self.min_value, self.max_value)}
+        """
+        values_dict = super().get_value()
+        return {self.name: self.base**v for v in values_dict.values()}
+    
+    def to_list(self):
+        values_list = super().to_list()
+        return [{self.name: self.base**v[self.name]} for v in values_list]
+    
+
 class UniformIntegerParameterRange(Parameter):
     """
     Uniform randomly sampled integer Hyper-Parameter object.