def test_transaction_resume_1(self):
sim = LambdaSimulation(2, lambda i: i, lambda i, c: [0, 1, 2],
lambda i, c, a: cast(float, a))
working_learner = ModuloLearner()
broken_learner = BrokenLearner()
#the second Experiment shouldn't ever call broken_factory() because
#we're resuming from the first experiment's transaction.log
try:
first_result = Experiment(
[sim], [working_learner],
evaluation_task=OnlineOnPolicyEvalTask(False)).evaluate(
"coba/tests/.temp/transactions.log")
second_result = Experiment(
[sim], [broken_learner],
evaluation_task=OnlineOnPolicyEvalTask(False)).evaluate(
"coba/tests/.temp/transactions.log")
actual_learners = second_result.learners.to_tuples()
actual_environments = second_result.environments.to_tuples()
actual_interactions = second_result.interactions.to_tuples()
expected_learners = [(0, "Modulo", "Modulo(p=0)", '0')]
expected_environments = [(0, 'LambdaSimulation')]
expected_interactions = [(0, 0, 1, 0), (0, 0, 2, 1)]
except Exception as e:
raise
finally:
if Path('coba/tests/.temp/transactions.log').exists():
Path('coba/tests/.temp/transactions.log').unlink()
self.assertCountEqual(actual_learners, expected_learners)
self.assertCountEqual(actual_environments, expected_environments)
self.assertCountEqual(actual_interactions, expected_interactions)
def test_process_rewards_info_logs_kwargs_partial(self):
task = OnlineOnPolicyEvalTask(time=False)
learner = RecordingLearner(with_info=True, with_log=True)
interactions = [
SimulatedInteraction(1,[1,2,3],rewards=[7,8,9]),
SimulatedInteraction(2,[4,5,6],rewards=[4,5,6],letters=['d','e','f']),
SimulatedInteraction(3,[7,8,9],rewards=[1,2,3],letters=['g','h','i']),
]
task_results = list(task.process(learner, interactions))
expected_predict_calls = [(1,[1,2,3]),(2,[4,5,6]),(3,[7,8,9])]
expected_predict_returns = [([1,0,0],1),([0,1,0],2),([0,0,1],3)]
expected_learn_calls = [(1,1,7,1,1),(2,5,5,1,2),(3,9,3,1,3)]
expected_task_results = [
{"rewards":7,'learn':1,'predict':1},
{"rewards":5,'learn':2,'predict':2,'letters':'e'},
{"rewards":3,'learn':3,'predict':3,'letters':'i'}
]
self.assertEqual(expected_predict_calls, learner.predict_calls)
self.assertEqual(expected_predict_returns, learner.predict_returns)
self.assertEqual(expected_learn_calls, learner.learn_calls)
self.assertEqual(expected_task_results, task_results)
def __init__(
self,
environments: Sequence[Environment],
learners: Sequence[Learner],
learner_task: LearnerTask = SimpleLearnerTask(),
environment_task: EnvironmentTask = SimpleEnvironmentTask(),
evaluation_task: EvaluationTask = OnlineOnPolicyEvalTask()
) -> None:
"""Instantiate an Experiment.
Args:
environments: The collection of environments to use in the experiment.
learners: The collection of learners to use in the experiment.
learner_task: A task which describes a learner.
environment_task: A task which describes an environment.
evaluation_task: A task which evaluates a learner on an environment.
"""
self._environments = environments
self._learners = learners
self._learner_task = learner_task
self._environment_task = environment_task
self._evaluation_task = evaluation_task
self._processes: Optional[int] = None
self._maxchunksperchild: Optional[int] = None
self._chunk_by: Optional[str] = None
def test_time(self):
task = OnlineOnPolicyEvalTask(time=True)
learner = RecordingLearner()
interactions = [SimulatedInteraction(1,[1,2,3],rewards=[7,8,9])]
task_results = list(task.process(learner, interactions))
self.assertAlmostEqual(0, task_results[0]["predict_time"], places=2)
self.assertAlmostEqual(0, task_results[0]["learn_time" ], places=2)
def test_no_params(self):
sim1 = NoParamsEnvironment()
learner = NoParamsLearner()
experiment = Experiment([sim1], [learner],
evaluation_task=OnlineOnPolicyEvalTask(False))
result = experiment.evaluate()
actual_learners = result.learners.to_tuples()
actual_environments = result.environments.to_tuples()
actual_interactions = result.interactions.to_tuples()
expected_learners = [(0, 'NoParamsLearner', 'NoParamsLearner')]
expected_environments = [(0, 'NoParamsEnvironment')]
expected_interactions = [(0, 0, 1, 0), (0, 0, 2, 1)]
self.assertCountEqual(actual_learners, expected_learners)
self.assertCountEqual(actual_environments, expected_environments)
self.assertCountEqual(actual_interactions, expected_interactions)
def test_sim(self):
sim1 = LambdaSimulation(2, lambda i: i, lambda i, c: [0, 1, 2],
lambda i, c, a: cast(float, a))
learner = ModuloLearner()
experiment = Experiment([sim1], [learner],
evaluation_task=OnlineOnPolicyEvalTask(False))
result = experiment.evaluate()
actual_learners = result.learners.to_tuples()
actual_environments = result.environments.to_tuples()
actual_interactions = result.interactions.to_tuples()
expected_learners = [(0, "Modulo", "Modulo(p=0)", '0')]
expected_environments = [(0, 'LambdaSimulation')]
expected_interactions = [(0, 0, 1, 0), (0, 0, 2, 1)]
self.assertCountEqual(actual_learners, expected_learners)
self.assertCountEqual(actual_environments, expected_environments)
self.assertCountEqual(actual_interactions, expected_interactions)
def test_process_sparse_rewards_no_info_no_logs_no_kwargs(self):
task = OnlineOnPolicyEvalTask(time=False)
learner = RecordingLearner(with_info=False, with_log=False)
interactions = [
SimulatedInteraction({'c':1},[{'a':1},{'a':2}],rewards=[7,8]),
SimulatedInteraction({'c':2},[{'a':4},{'a':5}],rewards=[4,5]),
]
task_results = list(task.process(learner, interactions))
expected_predict_calls = [({'c':1},[{'a':1},{'a':2}]),({'c':2},[{'a':4},{'a':5}])]
expected_predict_returns = [[1,0],[0,1]]
expected_learn_calls = [({'c':1},{'a':1},7,1,None),({'c':2},{'a':5},5,1,None)]
expected_task_results = [{"rewards":7},{"rewards":5}]
self.assertEqual(expected_predict_calls, learner.predict_calls)
self.assertEqual(expected_predict_returns, learner.predict_returns)
self.assertEqual(expected_learn_calls, learner.learn_calls)
self.assertEqual(expected_task_results, task_results)
def test_process_reveals_rewards_no_info_no_logs_no_kwargs(self):
task = OnlineOnPolicyEvalTask(time=False)
learner = RecordingLearner(with_info=False, with_log=False)
interactions = [
SimulatedInteraction(1,[1,2,3],reveals=[7,8,9],rewards=[1,3,5]),
SimulatedInteraction(2,[4,5,6],reveals=[4,5,6],rewards=[2,4,6]),
SimulatedInteraction(3,[7,8,9],reveals=[1,2,3],rewards=[3,5,7]),
]
task_results = list(task.process(learner, interactions))
expected_predict_calls = [(1,[1,2,3]),(2,[4,5,6]),(3,[7,8,9])]
expected_predict_returns = [[1,0,0],[0,1,0],[0,0,1]]
expected_learn_calls = [(1,1,7,1,None),(2,5,5,1,None),(3,9,3,1,None)]
expected_task_results = [{"reveals":7,"rewards":1},{"reveals":5,"rewards":4}, {"reveals":3,"rewards":7}]
self.assertEqual(expected_predict_calls, learner.predict_calls)
self.assertEqual(expected_predict_returns, learner.predict_returns)
self.assertEqual(expected_learn_calls, learner.learn_calls)
self.assertEqual(expected_task_results, task_results)
def test_learners(self):
sim = LambdaSimulation(2, lambda i: i, lambda i, c: [0, 1, 2],
lambda i, c, a: cast(float, a))
learner1 = ModuloLearner("0") #type: ignore
learner2 = ModuloLearner("1") #type: ignore
experiment = Experiment([sim], [learner1, learner2],
evaluation_task=OnlineOnPolicyEvalTask(False))
actual_result = experiment.evaluate()
actual_learners = actual_result._learners.to_tuples()
actual_environments = actual_result._environments.to_tuples()
actual_interactions = actual_result.interactions.to_tuples()
expected_learners = [(0, "Modulo", "Modulo(p=0)", '0'),
(1, "Modulo", "Modulo(p=1)", '1')]
expected_environments = [(0, 'LambdaSimulation')]
expected_interactions = [(0, 0, 1, 0), (0, 0, 2, 1), (0, 1, 1, 0),
(0, 1, 2, 1)]
self.assertCountEqual(actual_learners, expected_learners)
self.assertCountEqual(actual_environments, expected_environments)
self.assertCountEqual(actual_interactions, expected_interactions)
def test_ignore_raise(self):
CobaContext.logger = IndentLogger(ListSink())
sim1 = LambdaSimulation(2, lambda i: i, lambda i, c: [0, 1, 2],
lambda i, c, a: cast(float, a))
sim2 = LambdaSimulation(3, lambda i: i, lambda i, c: [3, 4, 5],
lambda i, c, a: cast(float, a))
experiment = Experiment(
[sim1, sim2], [ModuloLearner(), BrokenLearner()],
evaluation_task=OnlineOnPolicyEvalTask(False))
result = experiment.evaluate()
actual_learners = result.learners.to_tuples()
actual_environments = result.environments.to_tuples()
actual_interactions = result.interactions.to_tuples()
expected_learners = [(0, "Modulo", "Modulo(p=0)", '0'),
(1, "Broken", "Broken", float('nan'))]
expected_environments = [(0, 'LambdaSimulation'),
(1, 'LambdaSimulation')]
expected_interactions = [(0, 0, 1, 0), (0, 0, 2, 1), (1, 0, 1, 3),
(1, 0, 2, 4), (1, 0, 3, 5)]
self.assertIsInstance(CobaContext.logger, IndentLogger)
self.assertEqual(
2,
sum([
int("Unexpected exception:" in item)
for item in CobaContext.logger.sink.items
]))
self.assertCountEqual(actual_learners[0], expected_learners[0])
self.assertCountEqual(actual_learners[1][:3], expected_learners[1][:3])
self.assertTrue(math.isnan(expected_learners[1][3]))
self.assertCountEqual(actual_environments, expected_environments)
self.assertCountEqual(actual_interactions, expected_interactions)