def _clean_up_optimization():
"""Perform any cleanup necessary after completion of the optimization loop. Most notably,
this handles removal of temporary model files created for Keras optimization"""
for (root, dirs, files) in walk(TEMP_MODULES_DIR_PATH, topdown=False):
for file in files:
if file.startswith("__temp_"):
remove(f"{root}/{file}")
try:
rmdir(root)
except OSError:
G.warn_(f"Unidentified file found in temporary directory: {root}")
def _ask(self):
"""Suggest next point at which to evaluate the objective
Returns
-------
Some point in :attr:`space`, which is random while less than `n_initial_points` observations
have been `tell`-ed. After that, `base_estimator` is used to determine the next point
Notes
-----
If the suggested point has already been evaluated, a random point will be returned instead,
optionally accompanied by a warning message (depending on :attr:`warn_on_re_ask`)"""
if self._n_initial_points > 0 or self.base_estimator is None:
# Does not copy `self.rng` in order to keep advancing random state
return self.space.rvs(random_state=self.rng)[0]
else:
if not self.models:
raise RuntimeError(
"Random evaluations exhausted and no model has been fit")
#################### Check for Repeated Suggestion ####################
next_x = self._next_x
# Check distances between `next_x` and all evaluated points
min_delta_x = min(
[self.space.distance(next_x, xi) for xi in self.Xi])
if abs(min_delta_x) <= 1e-8: # `next_x` has already been evaluated
if self.warn_on_re_ask:
G.warn_("Repeated suggestion: {}".format(next_x))
# Set `_next_x` to random point, then re-invoke `_ask` to validate new point
self._next_x = self.space.rvs(random_state=self.rng)[0]
return self._ask()
# Return point computed from last call to `tell`
return next_x
def set_experiment_guidelines(
self,
model_initializer,
model_init_params,
model_extra_params=None,
feature_selector=None,
preprocessing_pipeline=None,
preprocessing_params=None,
notes=None,
do_raise_repeated=True,
):
"""Provide the arguments necessary to instantiate :class:`experiments.CrossValidationExperiment`. This method has the same
signature as :meth:`experiments.BaseExperiment.__init__` except where noted
Parameters
----------
model_initializer: Class, or functools.partial, or class instance
The algorithm class being used to initialize a model
model_init_params: Dict, or object
The dictionary of arguments given when creating a model instance with `model_initializer` via the `__init__` method
of :class:`models.Model`. Any kwargs that are considered valid by the `__init__` method of `model_initializer` are
valid in `model_init_params`
model_extra_params: Dict, or None, default=None
A dictionary of extra parameters passed to :class:`models.Model`. This is used to provide parameters to models'
non-initialization methods (like `fit`, `predict`, `predict_proba`, etc.), and for neural networks
feature_selector: List of str, callable, list of booleans, default=None
The value provided when splitting apart the input data for all provided DataFrames. `feature_selector` is provided as
the second argument for calls to `pandas.DataFrame.loc` in :meth:`BaseExperiment._initial_preprocessing`. If None,
`feature_selector` is set to all columns in :attr:`train_dataset`, less :attr:`target_column`, and :attr:`id_column`
preprocessing_pipeline: ...
... Experimental...
preprocessing_params: ...
... Experimental...
notes: String, or None, default=None
Additional information about the Experiment that will be saved with the Experiment's description result file. This
serves no purpose other than to facilitate saving Experiment details in a more readable format
do_raise_repeated: Boolean, default=False
If True and this Experiment locates a previous Experiment's results with matching Environment and Hyperparameter Keys,
a RepeatedExperimentError will be raised. Else, a warning will be logged
Notes
-----
The `auto_start` kwarg is not available here because :meth:`BaseOptimizationProtocol._execute_experiment` sets it to False
in order to check for duplicated keys before running the whole Experiment. This is the most notable difference between
calling :meth:`set_experiment_guidelines` and instantiating :class:`experiments.CrossValidationExperiment`"""
self.model_initializer = model_initializer
self.model_init_params = identify_algorithm_hyperparameters(
self.model_initializer)
try:
self.model_init_params.update(model_init_params)
except TypeError:
self.model_init_params.update(dict(build_fn=model_init_params))
self.model_extra_params = model_extra_params
self.feature_selector = feature_selector
self.preprocessing_pipeline = preprocessing_pipeline
self.preprocessing_params = preprocessing_params
self.notes = notes
self.do_raise_repeated = do_raise_repeated
if self.do_raise_repeated is False:
G.warn_(
'WARNING: Setting `do_raise_repeated`=False will allow Experiments to be unnecessarily duplicated'
)
self.algorithm_name, self.module_name = identify_algorithm(
self.model_initializer)
self._validate_guidelines()
#################### Deal with Keras ####################
if self.module_name == 'keras':
reusable_build_fn, reusable_wrapper_params, dummy_layers, dummy_compile_params = keras_prep_workflow(
self.model_initializer, self.model_init_params['build_fn'],
self.model_extra_params, self.source_script)
self.model_init_params = dict(build_fn=reusable_build_fn)
self.model_extra_params = reusable_wrapper_params
self.dummy_layers = dummy_layers
self.dummy_compile_params = dummy_compile_params
# FLAG: Deal with capitalization conflicts when comparing similar experiments: `optimizer`='Adam' vs 'adam'
self.set_dimensions()
