def test_reward_decreases_vail(demo_to_buffer: Any, use_actions: bool,
behavior_spec: BehaviorSpec, seed: int) -> None:
np.random.seed(seed)
torch.manual_seed(seed)
buffer_expert = create_agent_buffer(behavior_spec, 1000)
buffer_policy = create_agent_buffer(behavior_spec, 1000)
demo_to_buffer.return_value = None, buffer_expert
gail_settings = GAILSettings(demo_path="",
learning_rate=0.005,
use_vail=True,
use_actions=use_actions)
DiscriminatorNetwork.initial_beta = 0.0
# we must set the initial value of beta to 0 for testing
# If we do not, the kl-loss will dominate early and will block the estimator
gail_rp = create_reward_provider(RewardSignalType.GAIL, behavior_spec,
gail_settings)
for _ in range(300):
gail_rp.update(buffer_policy)
reward_expert = gail_rp.evaluate(buffer_expert)[0]
reward_policy = gail_rp.evaluate(buffer_policy)[0]
assert reward_expert >= 0 # GAIL / VAIL reward always positive
assert reward_policy >= 0
reward_expert = gail_rp.evaluate(buffer_expert)[0]
reward_policy = gail_rp.evaluate(buffer_policy)[0]
assert reward_expert > reward_policy # Expert reward greater than non-expert reward
def test_reward_decreases(demo_to_buffer: Any, use_actions: bool,
behavior_spec: BehaviorSpec, seed: int) -> None:
np.random.seed(seed)
torch.manual_seed(seed)
buffer_expert = create_agent_buffer(behavior_spec, 1000)
buffer_policy = create_agent_buffer(behavior_spec, 1000)
demo_to_buffer.return_value = None, buffer_expert
gail_settings = GAILSettings(demo_path="",
learning_rate=0.005,
use_vail=False,
use_actions=use_actions)
gail_rp = create_reward_provider(RewardSignalType.GAIL, behavior_spec,
gail_settings)
init_reward_expert = gail_rp.evaluate(buffer_expert)[0]
init_reward_policy = gail_rp.evaluate(buffer_policy)[0]
for _ in range(10):
gail_rp.update(buffer_policy)
reward_expert = gail_rp.evaluate(buffer_expert)[0]
reward_policy = gail_rp.evaluate(buffer_policy)[0]
assert reward_expert >= 0 # GAIL / VAIL reward always positive
assert reward_policy >= 0
reward_expert = gail_rp.evaluate(buffer_expert)[0]
reward_policy = gail_rp.evaluate(buffer_policy)[0]
assert reward_expert > reward_policy # Expert reward greater than non-expert reward
assert (reward_expert > init_reward_expert
) # Expert reward getting better as network trains
assert (reward_policy < init_reward_policy
) # Non-expert reward getting worse as network trains
def test_reward_provider_save(tmp_path, optimizer):
OptimizerClass, HyperparametersClass = optimizer
trainer_settings = TrainerSettings()
trainer_settings.hyperparameters = HyperparametersClass()
trainer_settings.reward_signals = {
RewardSignalType.CURIOSITY: CuriositySettings(),
RewardSignalType.GAIL: GAILSettings(demo_path=DEMO_PATH),
RewardSignalType.RND: RNDSettings(),
}
policy = create_policy_mock(trainer_settings, use_discrete=False)
optimizer = OptimizerClass(policy, trainer_settings)
# save at path 1
path1 = os.path.join(tmp_path, "runid1")
model_saver = TorchModelSaver(trainer_settings, path1)
model_saver.register(policy)
model_saver.register(optimizer)
model_saver.initialize_or_load()
policy.set_step(2000)
model_saver.save_checkpoint("MockBrain", 2000)
# create a new optimizer and policy
optimizer2 = OptimizerClass(policy, trainer_settings)
policy2 = create_policy_mock(trainer_settings, use_discrete=False)
# load weights
model_saver2 = TorchModelSaver(trainer_settings, path1, load=True)
model_saver2.register(policy2)
model_saver2.register(optimizer2)
model_saver2.initialize_or_load() # This is to load the optimizers
# assert the models have the same weights
module_dict_1 = optimizer.get_modules()
module_dict_2 = optimizer2.get_modules()
assert "Module:GAIL" in module_dict_1
assert "Module:GAIL" in module_dict_2
assert "Module:Curiosity" in module_dict_1
assert "Module:Curiosity" in module_dict_2
assert "Module:RND-pred" in module_dict_1
assert "Module:RND-pred" in module_dict_2
assert "Module:RND-target" in module_dict_1
assert "Module:RND-target" in module_dict_2
for name, module1 in module_dict_1.items():
assert name in module_dict_2
module2 = module_dict_2[name]
if hasattr(module1, "parameters"):
for param1, param2 in zip(module1.parameters(),
module2.parameters()):
assert param1.data.ne(param2.data).sum() == 0
# Run some rewards
data = create_agent_buffer(policy.behavior_spec, 1)
for reward_name in optimizer.reward_signals.keys():
rp_1 = optimizer.reward_signals[reward_name]
rp_2 = optimizer2.reward_signals[reward_name]
assert np.array_equal(rp_1.evaluate(data), rp_2.evaluate(data))
def test_reward_decreases(behavior_spec: BehaviorSpec, seed: int) -> None:
np.random.seed(seed)
torch.manual_seed(seed)
rnd_settings = RNDSettings(32, 0.01)
rnd_rp = RNDRewardProvider(behavior_spec, rnd_settings)
buffer = create_agent_buffer(behavior_spec, 5)
rnd_rp.update(buffer)
reward_old = rnd_rp.evaluate(buffer)[0]
for _ in range(100):
rnd_rp.update(buffer)
reward_new = rnd_rp.evaluate(buffer)[0]
assert reward_new < reward_old
def test_continuous_action_prediction(behavior_spec: BehaviorSpec, seed: int) -> None:
np.random.seed(seed)
torch.manual_seed(seed)
curiosity_settings = CuriositySettings(32, 0.1)
curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
buffer = create_agent_buffer(behavior_spec, 5)
for _ in range(200):
curiosity_rp.update(buffer)
prediction = curiosity_rp._network.predict_action(buffer)[0]
target = torch.tensor(buffer["actions"][0])
error = torch.mean((prediction - target) ** 2).item()
assert error < 0.001
def test_next_state_prediction(behavior_spec: BehaviorSpec, seed: int) -> None:
np.random.seed(seed)
torch.manual_seed(seed)
curiosity_settings = CuriositySettings(32, 0.1)
curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
buffer = create_agent_buffer(behavior_spec, 5)
for _ in range(100):
curiosity_rp.update(buffer)
prediction = curiosity_rp._network.predict_next_state(buffer)[0]
target = curiosity_rp._network.get_next_state(buffer)[0]
error = float(ModelUtils.to_numpy(torch.mean((prediction - target) ** 2)))
assert error < 0.001
def test_reward_decreases(behavior_spec: BehaviorSpec, seed: int) -> None:
np.random.seed(seed)
torch.manual_seed(seed)
curiosity_settings = CuriositySettings(32, 0.01)
curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
buffer = create_agent_buffer(behavior_spec, 5)
curiosity_rp.update(buffer)
reward_old = curiosity_rp.evaluate(buffer)[0]
for _ in range(20):
curiosity_rp.update(buffer)
reward_new = curiosity_rp.evaluate(buffer)[0]
assert reward_new < reward_old
def test_reward(behavior_spec: BehaviorSpec, reward: float) -> None:
buffer = create_agent_buffer(behavior_spec, 1000, reward)
settings = RewardSignalSettings()
extrinsic_rp = ExtrinsicRewardProvider(behavior_spec, settings)
generated_rewards = extrinsic_rp.evaluate(buffer)
assert (generated_rewards == reward).all()
# Test group rewards. Rewards should be double of the environment rewards, but shouldn't count
# the groupmate rewards.
buffer[BufferKey.GROUP_REWARD] = buffer[BufferKey.ENVIRONMENT_REWARDS]
# 2 agents with identical rewards
buffer[BufferKey.GROUPMATE_REWARDS].set(
[np.ones(1, dtype=np.float32) * reward] * 2
for _ in range(buffer.num_experiences))
generated_rewards = extrinsic_rp.evaluate(buffer)
assert (generated_rewards == 2 * reward).all()
# Test groupmate rewards. Total reward should be indiv_reward + 2 * teammate_reward + group_reward
extrinsic_rp = ExtrinsicRewardProvider(behavior_spec, settings)
extrinsic_rp.add_groupmate_rewards = True
generated_rewards = extrinsic_rp.evaluate(buffer)
assert (generated_rewards == 4 * reward).all()
def test_reward(behavior_spec: BehaviorSpec, reward: float) -> None:
buffer = create_agent_buffer(behavior_spec, 1000, reward)
settings = RewardSignalSettings()
extrinsic_rp = ExtrinsicRewardProvider(behavior_spec, settings)
generated_rewards = extrinsic_rp.evaluate(buffer)
assert (generated_rewards == reward).all()
