def test_curriculum_raises_all_completion_criteria_conversion():
with pytest.warns(TrainerConfigWarning):
run_options = RunOptions.from_dict(
yaml.safe_load(test_bad_curriculum_all_competion_criteria_config_yaml)
)
param_manager = EnvironmentParameterManager(
run_options.environment_parameters, 1337, False
)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (True, True)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (True, True)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (False, False)
assert param_manager.get_current_lesson_number() == {"param_1": 2}
def test_create_manager():
run_options = RunOptions.from_dict(yaml.safe_load(test_everything_config_yaml))
param_manager = EnvironmentParameterManager(
run_options.environment_parameters, 1337, False
)
assert param_manager.get_minimum_reward_buffer_size("fake_behavior") == 100
assert param_manager.get_current_lesson_number() == {
"param_1": 0,
"param_2": 0,
"param_3": 0,
}
assert param_manager.get_current_samplers() == {
"param_1": ConstantSettings(seed=1337, value=1),
"param_2": GaussianSettings(seed=1337 + 3, mean=4, st_dev=5),
"param_3": ConstantSettings(seed=1337 + 3 + 1, value=20),
}
# Not enough episodes completed
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 99},
) == (False, False)
# Not enough episodes reward
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1] * 101},
) == (False, False)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (True, True)
assert param_manager.get_current_lesson_number() == {
"param_1": 1,
"param_2": 0,
"param_3": 0,
}
param_manager_2 = EnvironmentParameterManager(
run_options.environment_parameters, 1337, restore=True
)
# The use of global status should make it so that the lesson numbers are maintained
assert param_manager_2.get_current_lesson_number() == {
"param_1": 1,
"param_2": 0,
"param_3": 0,
}
# No reset required
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 700},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [0] * 101},
) == (True, False)
assert param_manager.get_current_samplers() == {
"param_1": UniformSettings(seed=1337 + 2, min_value=1, max_value=3),
"param_2": GaussianSettings(seed=1337 + 3, mean=4, st_dev=5),
"param_3": ConstantSettings(seed=1337 + 3 + 1, value=20),
}
