def test_policy_size_shrink(self, context: Context) -> None:
"""Test limiting policy size over runs."""
# The main difference with the similar TD test is in reward signal propagated.
predictor = MCTS()
predictor._policy = {
("numpy", "2.0.0", "https://pypi.org/simple"): [1.0, 100],
("tensorflow", "2.0.0", "https://thoth-station.ninja/simple"):
[3.0, 100],
}
rewarded = list(predictor._policy.keys())[0] # numpy
state = flexmock(score=0.5)
state.should_receive(
"iter_resolved_dependencies").with_args().and_return([rewarded
]).once()
# No shrink as we are in this iteration.
context.iteration = 3 * mcts_module._MCTS_POLICY_SIZE_CHECK_ITERATION
old_policy_size = mcts_module._MCTS_POLICY_SIZE
with predictor.assigned_context(context):
try:
mcts_module._MCTS_POLICY_SIZE = 1
predictor.set_reward_signal(state, rewarded, math.inf)
finally:
mcts_module._MCTS_POLICY_SIZE = old_policy_size
# the numpy entry with a value of [1.5, 101] gets removed
assert predictor._policy == {
("tensorflow", "2.0.0", "https://thoth-station.ninja/simple"):
[3.0, 100],
}
assert predictor._next_state is None
def test_run_no_next_state(self, context: Context) -> None:
"""Test running the predictor when no next state is scheduled."""
predictor = MCTS()
assert predictor._next_state is None
# If no next state kept, we follow logic from the TD-learning.
flexmock(TemporalDifference)
state = flexmock()
unresolved_dependency = flexmock()
TemporalDifference.should_receive("run").with_args().and_return(state, unresolved_dependency).once()
context.iteration = 1000000 # Some big number not to hit the heat-up part.
with predictor.assigned_context(context):
assert predictor.run() == (state, unresolved_dependency)
def test_run_next_state(self, context: Context) -> None:
"""Test running the predictor when the next state is scheduled."""
state = flexmock()
unresolved_dependency = ("tensorflow", "2.0.0", "https://pypi.org/simple")
state.should_receive("get_random_unresolved_dependency").with_args(prefer_recent=True).and_return(
unresolved_dependency
).once()
predictor = MCTS()
predictor._next_state = state
context.beam.should_receive("get_last").and_return(state).once()
context.iteration = 1000000 # Some big number not to hit the heat-up part.
with predictor.assigned_context(context):
assert predictor.run() == (state, unresolved_dependency)
def test_run_heat_up(self, context: Context, next_state) -> None:
"""Test running the predictor in the "heat-up" phase regardless next state being set."""
state = flexmock()
unresolved_dependency = ("tensorflow", "2.0.0", "https://pypi.org/simple")
predictor = MCTS()
predictor._next_state = None
flexmock(TemporalDifference)
TemporalDifference.should_receive("run").with_args().and_return(state, unresolved_dependency).once()
context.iteration = 1 # Some small number to hit the heat-up part.
with predictor.assigned_context(context):
assert predictor.run() == (state, unresolved_dependency)
def test_run_next_state_no_last(self, context: Context) -> None:
"""Test running the predictor when the next state is not last state added to beam."""
state = flexmock()
unresolved_dependency = ("tensorflow", "2.0.0", "https://pypi.org/simple")
predictor = MCTS()
predictor._next_state = flexmock()
context.beam.should_receive("get_last").and_return(flexmock()).once()
flexmock(TemporalDifference)
TemporalDifference.should_receive("run").with_args().and_return(state, unresolved_dependency).once()
context.iteration = 1000000 # Some big number not to hit the heat-up part.
with predictor.assigned_context(context):
assert predictor.run() == (state, unresolved_dependency)