def test_set_reward_signal_nan_inf(self, float_case: float) -> None:
"""Test (not) keeping the reward signal for nan/inf."""
predictor = TemporalDifference()
state = State()
state.add_resolved_dependency(
("tensorflow", "2.3.0", "https://pypi.org/simple"))
state.add_resolved_dependency(
("flask", "0.12", "https://pypi.org/simple"))
state.add_unresolved_dependency(
("termial-random", "0.0.2", "https://pypi.org/simple"))
predictor._policy = {
("flask", "0.12", "https://pypi.org/simple"): [0.2, 1],
}
predictor._steps_taken = 2
predictor._steps_reward = 1.2
predictor._next_state = state
assert (predictor.set_reward_signal(
state, ("tensorflow", "2.0.0", "https://pypi.org/simple"),
float_case) is None)
assert predictor._policy == {
("flask", "0.12", "https://pypi.org/simple"): [1.4, 2],
("tensorflow", "2.3.0", "https://pypi.org/simple"): [1.2, 1],
}
assert predictor._steps_taken == 0
assert predictor._steps_reward == 0.0
assert predictor._next_state is None
def test_n_step_td_step_no_adjust(self, context: Context) -> None:
"""Test adjusting steps taken on reward signal propagation."""
predictor = TemporalDifference(step=1)
predictor._temperature = 1.0
predictor._steps_taken = 0
package_tuple = ("tensorflow", "2.3.1", "https://pypi.org/simple")
state = State()
state.add_resolved_dependency(package_tuple)
with predictor.assigned_context(context):
predictor.set_reward_signal(state, package_tuple, 0.33)
assert predictor._policy.get(package_tuple) is None
predictor._steps_taken = 1
with predictor.assigned_context(context):
predictor.set_reward_signal(state, package_tuple, 0.2)
assert predictor._policy.get(package_tuple) == [0.53, 1]
assert predictor._steps_taken == 0
def test_set_reward_signal_unseen(self) -> None:
"""Test keeping the reward signal for an unseen step."""
reward = 42.24
package_tuple = ("tensorflow", "2.0.0", "https://thoth-station.ninja")
state = flexmock()
state.should_receive("iter_resolved_dependencies").and_return([package_tuple]).once()
predictor = TemporalDifference()
predictor._policy = {
("numpy", "1.0.0", "https://pypi.org/simple"): [30.30, 92],
}
predictor._steps_taken = 1
predictor.set_reward_signal(state, None, reward)
assert predictor._policy == {
package_tuple: [42.24, 1],
("numpy", "1.0.0", "https://pypi.org/simple"): [30.30, 92],
}
def test_run_exploration(self, context: Context) -> None:
"""Tests run when exploration is performed."""
flexmock(TemporalDifference)
flexmock(TemporalDifference)
flexmock(AdaptiveSimulatedAnnealing)
flexmock(State)
max_state = State(score=3.0)
probable_state = State(score=2.0)
context.beam.add_state(max_state)
context.beam.add_state(probable_state)
unresolved_dependency = (
"pytorch",
"1.0.0",
"https://thoth-station.ninja/simple",
)
flexmock(random)
random.should_receive("randrange").with_args(1, 2).and_return(0).once()
random.should_receive("random").and_return(0.50).once(
) # *lower* than acceptance_probability that is 0.75 so we do exploitation
probable_state.should_receive(
"get_random_unresolved_dependency").with_args(
prefer_recent=True).and_return(unresolved_dependency).once()
TemporalDifference.should_receive("_temperature_function").with_args(
1.0, context).and_return(0.9).once()
AdaptiveSimulatedAnnealing.should_receive(
"_compute_acceptance_probability").with_args(
max_state.score, probable_state.score,
0.9).and_return(0.75).once()
context.beam.should_receive("max").with_args().and_return(
max_state).once()
predictor = TemporalDifference(step=1)
predictor._steps_taken = 0
predictor._temperature = 1.0
with predictor.assigned_context(context):
assert predictor.run() == (probable_state, unresolved_dependency)
assert predictor._steps_taken == 1