def setup(self):
"""
Tests whether the initialization works correctly.
Tests:
- optimizer correct
- minimization correct
- param_defs correct
"""
optimizer = "RandomSearch"
name = "test_init_experiment"
self.param_defs = {
"x": MinMaxNumericParamDef(0, 1),
"name": NominalParamDef(["A", "B", "C"])
}
minimization = True
optimizer_params = {"multiprocessing": "none"}
exp = experiment.Experiment(name, self.param_defs, minimization)
self.EAss = ExperimentAssistant(optimizer,
exp,
optimizer_arguments=optimizer_params)
assert_equal(self.EAss._optimizer.__class__.__name__, optimizer)
assert_equal(self.EAss._optimizer_arguments, optimizer_params)
assert_equal(self.EAss._experiment.minimization_problem, minimization)
def init_experiment(self, name, optimizer, param_defs, exp_id=None,
notes=None, optimizer_arguments=None,
minimization=True):
"""
Initializes an experiment.
Parameters
----------
name : string
name of the experiment.
optimizer : string
String representation of the optimizer.
param_defs : dict of parameter definitions
Dictionary of parameter definition classes.
optimizer_arguments : dict, optional
A dictionary defining the operation of the optimizer. See the
respective documentation of the optimizers.
Default is None, which are default values.
exp_id : string or None, optional
The id of the experiment, which will be used to reference it.
Should be a proper uuid, and especially has to be unique. If it is
not, an error may be returned.
notes : jsonable object or None, optional
Any note that you'd like to put in the experiment. Could be used
to provide some details on the experiment, on the start time or the
user starting it.
minimization : bool, optional
Whether the problem is one of minimization. Defaults to True.
Returns
-------
exp_id : string
String representing the id of the experiment or "failed" if failed.
Raises
------
ValueError :
Iff there already is an experiment with the exp_id for this lab
assistant. Does not occur if no exp_id is given.
"""
if exp_id in self._exp_assistants.keys():
raise ValueError("Already an experiment with id %s registered."
%exp_id)
if exp_id is None:
while True:
exp_id = uuid.uuid4().hex
if exp_id not in self._exp_assistants.keys():
break
exp_ass = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_arguments,
write_directory_base=self._lab_run_directory,
csv_write_frequency=1)
exp_ass.init_experiment(name, param_defs, exp_id, notes, minimization)
self._exp_assistants[exp_id] = exp_ass
self._logger.info("Experiment initialized successfully.")
return exp_id
def init_experiment(self, name, optimizer, param_defs, exp_id=None,
notes=None, optimizer_arguments=None, minimization=True):
"""
Initializes a new experiment.
This actually initializes self.cv many experiments.
Internally, the experiments are called name_i, where name is the
experiment_name and i is the number of the experiment.
Parameters
----------
name : string
The name of the experiment. This has to be unique.
optimizer : Optimizer instance or string
This is an optimizer implementing the corresponding functions: It
gets an experiment instance, and returns one or multiple candidates
which should be evaluated next.
Alternatively, it can be a string corresponding to the optimizer,
as defined by apsis.utilities.optimizer_utils.
param_defs : dict of ParamDef.
This is the parameter space defining the experiment.
optimizer_arguments : dict, optional
These are arguments for the optimizer. Refer to their documentation
as to which are available.
minimization : bool, optional
Whether the problem is one of minimization or maximization.
"""
self._logger.info("Initializing new experiment \"%s\". "
" Parameter definitions: %s. Minimization is %s"
%(name, param_defs, minimization))
if exp_id in self._exp_assistants:
raise ValueError("Already an experiment with id %s registered."
%name)
if exp_id is None:
while True:
exp_id = uuid.uuid4().hex
if exp_id not in self._exp_assistants.keys():
break
self._exp_assistants[exp_id] = []
self.candidates_pending[exp_id] = []
for i in range(self.cv):
exp_ass = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_arguments,
write_directory_base=self._lab_run_directory,
csv_write_frequency=1)
exp_ass.init_experiment(name + "_" + str(i), param_defs, exp_id,
notes, minimization)
self._exp_assistants[exp_id].append(exp_ass)
self.candidates_pending[exp_id].append([])
self._logger.info("Experiment initialized successfully.")
return exp_id
def setup(self):
"""
Tests whether the initialization works correctly.
Tests:
- optimizer correct
- minimization correct
- param_defs correct
"""
optimizer = "RandomSearch"
name = "test_init_experiment"
self.param_defs = {
"x": MinMaxNumericParamDef(0, 1),
"name": NominalParamDef(["A", "B", "C"])
}
minimization = True
optimizer_params = {
"multiprocessing": "none"
}
self.EAss = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_params)
self.EAss.init_experiment(name, param_defs=self.param_defs, minimization=minimization)
assert_equal(self.EAss._optimizer.__class__.__name__, optimizer)
assert_equal(self.EAss._optimizer_arguments, optimizer_params)
assert_equal(self.EAss._experiment.minimization_problem, minimization)
def test_init_experiment(self):
optimizer = "RandomSearch"
name = "test_init_experiment"
self.param_defs = {
"x": MinMaxNumericParamDef(0, 1),
"name": NominalParamDef(["A", "B", "C"])
}
minimization = True
optimizer_params = {
"multiprocessing": "none"
}
self.EAss = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_params)
self.EAss.init_experiment(name, param_defs=self.param_defs, minimization=minimization)
with assert_raises(ValueError):
self.EAss.init_experiment(name, param_defs=self.param_defs, minimization=minimization)
with assert_raises(ValueError):
self.EAss.set_experiment("this value does not matter.")
def _load_exp_assistant_from_path(self, path):
"""
This loads a complete exp_assistant from path.
Specifically, it looks for exp_assistant.json in the path and restores
optimizer_class, optimizer_arguments and write_dir from this. It then
loads the experiment from the write_dir/experiment.json, then
initializes both.
Parameters
----------
path : string
The path from which to initialize. This must contain an
exp_assistant.json as specified.
"""
self._logger.debug("Loading Exp_assistant from path %s" % path)
with open(path + "/exp_assistant.json", 'r') as infile:
exp_assistant_json = json.load(infile)
optimizer_class = exp_assistant_json["optimizer_class"]
optimizer_arguments = exp_assistant_json["optimizer_arguments"]
exp_ass_write_dir = exp_assistant_json["write_dir"]
ensure_directory_exists(exp_ass_write_dir)
self._logger.debug(
"\tLoaded exp_parameters: "
"optimizer_class: %s, optimizer_arguments: %s,"
"write_dir: %s" %
(optimizer_class, optimizer_arguments, exp_ass_write_dir))
exp = self._load_experiment(path)
self._logger.debug("\tLoaded Experiment. %s" % exp.to_dict())
exp_ass = ExperimentAssistant(optimizer_class=optimizer_class,
experiment=exp,
optimizer_arguments=optimizer_arguments,
write_dir=exp_ass_write_dir)
if exp_ass.exp_id in self._exp_assistants:
raise ValueError("Loaded exp_id is duplicated in experiment! id "
"is %s" % exp_ass.exp_id)
self._exp_assistants[exp_ass.exp_id] = exp_ass
self._logger.info("Successfully loaded experiment from %s." % path)
class TestExperimentAssistant(object):
"""
Tests the experiment assistant.
"""
EAss = None
param_defs = None
def setup(self):
"""
Tests whether the initialization works correctly.
Tests:
- optimizer correct
- minimization correct
- param_defs correct
"""
optimizer = "RandomSearch"
name = "test_init_experiment"
self.param_defs = {
"x": MinMaxNumericParamDef(0, 1),
"name": NominalParamDef(["A", "B", "C"])
}
minimization = True
optimizer_params = {
"multiprocessing": "none"
}
self.EAss = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_params)
self.EAss.init_experiment(name, param_defs=self.param_defs, minimization=minimization)
assert_equal(self.EAss._optimizer.__class__.__name__, optimizer)
assert_equal(self.EAss._optimizer_arguments, optimizer_params)
assert_equal(self.EAss._experiment.minimization_problem, minimization)
def test_init_experiment(self):
optimizer = "RandomSearch"
name = "test_init_experiment"
self.param_defs = {
"x": MinMaxNumericParamDef(0, 1),
"name": NominalParamDef(["A", "B", "C"])
}
minimization = True
optimizer_params = {
"multiprocessing": "none"
}
self.EAss = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_params)
self.EAss.init_experiment(name, param_defs=self.param_defs, minimization=minimization)
with assert_raises(ValueError):
self.EAss.init_experiment(name, param_defs=self.param_defs, minimization=minimization)
with assert_raises(ValueError):
self.EAss.set_experiment("this value does not matter.")
def teardown(self):
self.EAss.set_exit()
def test_get_next_candidate(self):
"""
Tests the get next candidate function.
Tests:
- The candidate's parameters are acceptable
"""
cand = None
counter = 0
while cand is None and counter < 20:
cand = self.EAss.get_next_candidate()
time.sleep(0.1)
counter += 1
if counter == 20:
raise Exception("Received no result in the first 2 seconds.")
assert_is_none(cand.result)
params = cand.params
assert_less_equal(params["x"], 1)
assert_greater_equal(params["x"], 0)
assert_in(params["name"], self.param_defs["name"].values)
self.EAss.update(cand, "pausing")
time.sleep(1)
new_cand = None
while new_cand is None and counter < 20:
new_cand = self.EAss.get_next_candidate()
time.sleep(0.1)
counter += 1
if counter == 20:
raise Exception("Received no result in the first 2 seconds.")
assert_equal(new_cand, cand)
def test_update(self):
"""
Tests whether update works.
- candidate exists in the list
- result is equal
- the status message incorrect error works
- the candidate instance check works
"""
cand = self.EAss.get_next_candidate()
cand.result = 1
self.EAss.update(cand)
assert_items_equal(self.EAss._experiment.candidates_finished, [cand])
assert_equal(self.EAss._experiment.candidates_finished[0].result, 1)
self.EAss.update(cand, "pausing")
self.EAss.update(cand, "working")
with assert_raises(ValueError):
self.EAss.update(cand, status="No status.")
with assert_raises(ValueError):
self.EAss.update(False)
def test_get_best_candidate(self):
"""
Tests whether get_best_candidate works.
- Whether the best of the two candidates is the one it should be.
"""
cand_one = self.EAss.get_next_candidate()
cand_one.result = 1
self.EAss.update(cand_one)
cand_two = self.EAss.get_next_candidate()
cand_two.result = 0
self.EAss.update(cand_two)
assert_equal(cand_two, self.EAss.get_best_candidate())
def test_all_plots_working(self):
"""
Tests whether all of the plot functions work. Does not test for correctness.
"""
cand = self.EAss.get_next_candidate()
cand.result = 1
self.EAss.update(cand)
cand = self.EAss.get_next_candidate()
cand.result = 0
cand = self.EAss.get_next_candidate()
cand.result = 2
self.EAss.plot_result_per_step()
def test_get_candidates_dict(self):
candidates_dict = self.EAss.get_candidates()
assert_true(isinstance(candidates_dict, dict))
for l in ["finished", "pending", "working"]:
assert_in(l, candidates_dict)
assert_true(isinstance(candidates_dict[l], list))
def init_experiment(self,
name,
optimizer,
param_defs,
exp_id=None,
notes=None,
optimizer_arguments=None,
minimization=True):
"""
Initializes an experiment.
Parameters
----------
name : string
name of the experiment.
optimizer : string
String representation of the optimizer.
param_defs : dict of parameter definitions
Dictionary of parameter definition classes.
optimizer_arguments : dict, optional
A dictionary defining the operation of the optimizer. See the
respective documentation of the optimizers.
Default is None, which are default values.
exp_id : string or None, optional
The id of the experiment, which will be used to reference it.
Should be a proper uuid, and especially has to be unique. If it is
not, an error may be returned.
notes : jsonable object or None, optional
Any note that you'd like to put in the experiment. Could be used
to provide some details on the experiment, on the start time or the
user starting it.
minimization : bool, optional
Whether the problem is one of minimization. Defaults to True.
Returns
-------
exp_id : string
String representing the id of the experiment or "failed" if failed.
Raises
------
ValueError :
Iff there already is an experiment with the exp_id for this lab
assistant. Does not occur if no exp_id is given.
"""
self._logger.debug("Initializing new experiment. Parameters: "
"name: %s, optimizer: %s, param_defs: %s, "
"exp_id: %s, notes: %s, optimizer_arguments: %s, "
"minimization: %s" %
(name, optimizer, param_defs, exp_id, notes,
optimizer_arguments, minimization))
if exp_id in self._exp_assistants.keys():
raise ValueError("Already an experiment with id %s registered." %
exp_id)
if exp_id is None:
while True:
exp_id = uuid.uuid4().hex
if exp_id not in self._exp_assistants.keys():
break
self._logger.debug("\tGenerated new exp_id: %s" % exp_id)
if not self._write_dir:
exp_assistant_write_directory = None
else:
exp_assistant_write_directory = os.path.join(self._write_dir +
"/" + exp_id)
ensure_directory_exists(exp_assistant_write_directory)
self._logger.debug("\tExp_ass directory: %s" %
exp_assistant_write_directory)
exp = experiment.Experiment(name, param_defs, exp_id, notes,
minimization)
exp_ass = ExperimentAssistant(optimizer,
experiment=exp,
optimizer_arguments=optimizer_arguments,
write_dir=exp_assistant_write_directory)
self._exp_assistants[exp_id] = exp_ass
self._logger.info("Experiment initialized successfully with id %s." %
exp_id)
self._write_state_to_file()
return exp_id
def clone_experiments_by_id(self, exp_id, optimizer,
optimizer_arguments, new_exp_name):
"""
Take an existing experiment managed by this lab assistant,
fully clone it and store it under a new name to use it with a new
optimizer. This functionality can be used to initialize several experiments
of several optimizers with the same points.
For the given exp_name all underlying experiment instances are cloned and renamed.
Then a new experiment assistant is instantiated given the cloned and renamed
experiment using the given optimizer. The new experiment assistants are stored
and managed inside this lab assistant. The old experiment is not touched
and continues to be part of this lab assistant.
The parameter definitions and other experiment specific configuration is
copied over from the old to the new experiment.
Parameters
----------
exp_id : string
The id of the experiment to be cloned.
new_exp_name: string, optional
The name the cloned experiment will have after creation. If None,
the old name is reused.
optimizer : Optimizer instance or string
This is an optimizer implementing the corresponding functions: It
gets an experiment instance, and returns one or multiple candidates
which should be evaluated next.
Alternatively, it can be a string corresponding to the optimizer,
as defined by apsis.utilities.optimizer_utils.
optimizer_arguments : dict, optional
These are arguments for the optimizer. Refer to their documentation
as to which are available.
"""
while True:
new_exp_id = uuid.uuid4().hex
if new_exp_id not in self._exp_assistants.keys():
break
self._exp_assistants[new_exp_id] = []
#every experiment has self.cv many assistants
for i in range(len(self._exp_assistants[exp_id])):
old_exp_assistant = self._exp_assistants[exp_id][i]
old_exp = old_exp_assistant._experiment
#clone and rename experiment
new_exp = old_exp.clone()
new_name_cved = new_exp_name + "_" + str(i)
new_exp.name = new_name_cved
#recreate exp assistant
new_exp_assistant = ExperimentAssistant(optimizer, optimizer_arguments=optimizer_arguments,
write_directory_base=self._lab_run_directory,
csv_write_frequency=1)
new_exp_assistant.set_experiment(new_exp)
self._exp_assistants[new_exp_id].append(new_exp_assistant)
self.candidates_pending[new_exp_id] = copy.deepcopy(self.candidates_pending[exp_id])
self._logger.info("Experiment " + str(exp_id) + " cloned to " +
str(new_exp_id) + " and successfully initialized.")
return new_exp_id
class TestExperimentAssistant(object):
"""
Tests the experiment assistant.
"""
EAss = None
param_defs = None
def setup(self):
"""
Tests whether the initialization works correctly.
Tests:
- optimizer correct
- minimization correct
- param_defs correct
"""
optimizer = "RandomSearch"
name = "test_init_experiment"
self.param_defs = {
"x": MinMaxNumericParamDef(0, 1),
"name": NominalParamDef(["A", "B", "C"])
}
minimization = True
optimizer_params = {"multiprocessing": "none"}
exp = experiment.Experiment(name, self.param_defs, minimization)
self.EAss = ExperimentAssistant(optimizer,
exp,
optimizer_arguments=optimizer_params)
assert_equal(self.EAss._optimizer.__class__.__name__, optimizer)
assert_equal(self.EAss._optimizer_arguments, optimizer_params)
assert_equal(self.EAss._experiment.minimization_problem, minimization)
def teardown(self):
self.EAss.set_exit()
def test_get_next_candidate(self):
"""
Tests the get next candidate function.
Tests:
- The candidate's parameters are acceptable
"""
cand = None
counter = 0
while cand is None and counter < 20:
cand = self.EAss.get_next_candidate()
time.sleep(0.1)
counter += 1
if counter == 20:
raise Exception("Received no result in the first 2 seconds.")
assert_is_none(cand.result)
params = cand.params
assert_less_equal(params["x"], 1)
assert_greater_equal(params["x"], 0)
assert_in(params["name"], self.param_defs["name"].values)
self.EAss.update(cand, "pausing")
time.sleep(1)
new_cand = None
while new_cand is None and counter < 20:
new_cand = self.EAss.get_next_candidate()
time.sleep(0.1)
counter += 1
if counter == 20:
raise Exception("Received no result in the first 2 seconds.")
assert_equal(new_cand, cand)
def test_update(self):
"""
Tests whether update works.
- candidate exists in the list
- result is equal
- the status message incorrect error works
- the candidate instance check works
"""
cand = self.EAss.get_next_candidate()
cand.result = 1
self.EAss.update(cand)
assert_items_equal(self.EAss._experiment.candidates_finished, [cand])
assert_equal(self.EAss._experiment.candidates_finished[0].result, 1)
self.EAss.update(cand, "pausing")
self.EAss.update(cand, "working")
with assert_raises(ValueError):
self.EAss.update(cand, status="No status.")
with assert_raises(ValueError):
self.EAss.update(False)
def test_get_best_candidate(self):
"""
Tests whether get_best_candidate works.
- Whether the best of the two candidates is the one it should be.
"""
cand_one = self.EAss.get_next_candidate()
cand_one.result = 1
self.EAss.update(cand_one)
cand_two = self.EAss.get_next_candidate()
cand_two.result = 0
self.EAss.update(cand_two)
assert_equal(cand_two, self.EAss.get_best_candidate())
def test_all_plots_working(self):
"""
Tests whether all of the plot functions work. Does not test for correctness.
"""
cand = self.EAss.get_next_candidate()
cand.result = 1
self.EAss.update(cand)
cand = self.EAss.get_next_candidate()
cand.result = 0
cand = self.EAss.get_next_candidate()
cand.result = 2
self.EAss.plot_result_per_step()
def test_get_candidates_dict(self):
candidates_dict = self.EAss.get_candidates()
assert_true(isinstance(candidates_dict, dict))
for l in ["finished", "pending", "working"]:
assert_in(l, candidates_dict)
assert_true(isinstance(candidates_dict[l], list))
