def test_networkbody_lstm():
torch.manual_seed(0)
obs_size = 4
seq_len = 6
network_settings = NetworkSettings(memory=NetworkSettings.MemorySettings(
sequence_length=seq_len, memory_size=12))
obs_shapes = [(obs_size, )]
networkbody = NetworkBody(create_observation_specs_with_shapes(obs_shapes),
network_settings)
optimizer = torch.optim.Adam(networkbody.parameters(), lr=3e-4)
sample_obs = torch.ones((seq_len, obs_size))
for _ in range(300):
encoded, _ = networkbody([sample_obs],
memories=torch.ones(1, 1, 12),
sequence_length=seq_len)
# Try to force output to 1
loss = torch.nn.functional.mse_loss(encoded, torch.ones(encoded.shape))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# In the last step, values should be close to 1
for _enc in encoded.flatten().tolist():
assert _enc == pytest.approx(1.0, abs=0.1)
def test_actor_critic(ac_type, lstm):
obs_size = 4
network_settings = NetworkSettings(
memory=NetworkSettings.MemorySettings() if lstm else None,
normalize=True)
obs_spec = create_observation_specs_with_shapes([(obs_size, )])
act_size = 2
mask = torch.ones([1, act_size * 2])
stream_names = [f"stream_name{n}" for n in range(4)]
# action_spec = ActionSpec.create_continuous(act_size[0])
action_spec = ActionSpec(act_size,
tuple(act_size for _ in range(act_size)))
actor = ac_type(obs_spec, network_settings, action_spec, stream_names)
if lstm:
sample_obs = torch.ones(
(1, network_settings.memory.sequence_length, obs_size))
memories = torch.ones(
(1, network_settings.memory.sequence_length, actor.memory_size))
else:
sample_obs = torch.ones((1, obs_size))
memories = torch.tensor([])
# memories isn't always set to None, the network should be able to
# deal with that.
# Test critic pass
value_out, memories_out = actor.critic_pass([sample_obs],
memories=memories)
for stream in stream_names:
if lstm:
assert value_out[stream].shape == (
network_settings.memory.sequence_length, )
assert memories_out.shape == memories.shape
else:
assert value_out[stream].shape == (1, )
# Test get action stats and_value
action, log_probs, entropies, value_out, mem_out = actor.get_action_stats_and_value(
[sample_obs], memories=memories, masks=mask)
if lstm:
assert action.continuous_tensor.shape == (64, 2)
else:
assert action.continuous_tensor.shape == (1, 2)
assert len(action.discrete_list) == 2
for _disc in action.discrete_list:
if lstm:
assert _disc.shape == (64, 1)
else:
assert _disc.shape == (1, 1)
if mem_out is not None:
assert mem_out.shape == memories.shape
for stream in stream_names:
if lstm:
assert value_out[stream].shape == (
network_settings.memory.sequence_length, )
else:
assert value_out[stream].shape == (1, )
def test_bad_config():
brain_params = make_brain_parameters(discrete_action=False,
visual_inputs=0,
vec_obs_size=6)
# Test that we throw an error if we have sequence length greater than batch size
with pytest.raises(TrainerConfigError):
TrainerSettings(
network_settings=NetworkSettings(
memory=NetworkSettings.MemorySettings(sequence_length=64)),
hyperparameters=PPOSettings(batch_size=32),
)
_ = PPOTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
def test_multinetworkbody_lstm(with_actions):
torch.manual_seed(0)
obs_size = 4
act_size = 2
seq_len = 16
n_agents = 3
network_settings = NetworkSettings(memory=NetworkSettings.MemorySettings(
sequence_length=seq_len, memory_size=12))
obs_shapes = [(obs_size, )]
action_spec = ActionSpec(act_size,
tuple(act_size for _ in range(act_size)))
networkbody = MultiAgentNetworkBody(
create_observation_specs_with_shapes(obs_shapes), network_settings,
action_spec)
optimizer = torch.optim.Adam(networkbody.parameters(), lr=3e-4)
sample_obs = [[0.1 * torch.ones((seq_len, obs_size))]
for _ in range(n_agents)]
# simulate baseline in POCA
sample_act = [
AgentAction(
0.1 * torch.ones((seq_len, 2)),
[0.1 * torch.ones(seq_len) for _ in range(act_size)],
) for _ in range(n_agents - 1)
]
for _ in range(300):
if with_actions:
encoded, _ = networkbody(
obs_only=sample_obs[:1],
obs=sample_obs[1:],
actions=sample_act,
memories=torch.ones(1, 1, 12),
sequence_length=seq_len,
)
else:
encoded, _ = networkbody(
obs_only=sample_obs,
obs=[],
actions=[],
memories=torch.ones(1, 1, 12),
sequence_length=seq_len,
)
# Try to force output to 1
loss = torch.nn.functional.mse_loss(encoded, torch.ones(encoded.shape))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# In the last step, values should be close to 1
for _enc in encoded.flatten().tolist():
assert _enc == pytest.approx(1.0, abs=0.1)
def test_networkbody_visual():
torch.manual_seed(0)
vec_obs_size = 4
obs_size = (84, 84, 3)
network_settings = NetworkSettings()
obs_shapes = [(vec_obs_size, ), obs_size]
networkbody = NetworkBody(create_observation_specs_with_shapes(obs_shapes),
network_settings)
optimizer = torch.optim.Adam(networkbody.parameters(), lr=3e-3)
sample_obs = 0.1 * torch.ones((1, 84, 84, 3))
sample_vec_obs = torch.ones((1, vec_obs_size))
obs = [sample_vec_obs] + [sample_obs]
for _ in range(150):
encoded, _ = networkbody(obs)
assert encoded.shape == (1, network_settings.hidden_units)
# Try to force output to 1
loss = torch.nn.functional.mse_loss(encoded, torch.ones(encoded.shape))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# In the last step, values should be close to 1
for _enc in encoded.flatten().tolist():
assert _enc == pytest.approx(1.0, abs=0.1)
def create_bc_module(mock_behavior_specs, bc_settings, use_rnn, tanhresample):
# model_path = env.external_brain_names[0]
trainer_config = TrainerSettings()
trainer_config.network_settings.memory = (NetworkSettings.MemorySettings()
if use_rnn else None)
policy = NNPolicy(
0,
mock_behavior_specs,
trainer_config,
False,
"test",
False,
tanhresample,
tanhresample,
)
with policy.graph.as_default():
bc_module = BCModule(
policy,
policy_learning_rate=trainer_config.hyperparameters.learning_rate,
default_batch_size=trainer_config.hyperparameters.batch_size,
default_num_epoch=3,
settings=bc_settings,
)
policy.initialize_or_load(
) # Normally the optimizer calls this after the BCModule is created
return bc_module
def test_normalization():
behavior_spec = mb.setup_test_behavior_specs(use_discrete=True,
use_visual=False,
vector_action_space=[2],
vector_obs_space=1)
time_horizon = 6
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
observation_shapes=[(1, )],
action_space=[2],
)
# Change half of the obs to 0
for i in range(3):
trajectory.steps[i].obs[0] = np.zeros(1, dtype=np.float32)
policy = NNPolicy(
0,
behavior_spec,
TrainerSettings(network_settings=NetworkSettings(normalize=True)),
False,
"testdir",
False,
)
trajectory_buffer = trajectory.to_agentbuffer()
policy.update_normalization(trajectory_buffer["vector_obs"])
# Check that the running mean and variance is correct
steps, mean, variance = policy.sess.run([
policy.normalization_steps, policy.running_mean,
policy.running_variance
])
assert steps == 6
assert mean[0] == 0.5
# Note: variance is divided by number of steps, and initialized to 1 to avoid
# divide by 0. The right answer is 0.25
assert (variance[0] - 1) / steps == 0.25
# Make another update, this time with all 1's
time_horizon = 10
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
observation_shapes=[(1, )],
action_space=[2],
)
trajectory_buffer = trajectory.to_agentbuffer()
policy.update_normalization(trajectory_buffer["vector_obs"])
# Check that the running mean and variance is correct
steps, mean, variance = policy.sess.run([
policy.normalization_steps, policy.running_mean,
policy.running_variance
])
assert steps == 16
assert mean[0] == 0.8125
assert (variance[0] - 1) / steps == pytest.approx(0.152, abs=0.01)
def test_trainer_update_policy(
dummy_config, curiosity_dummy_config, use_discrete # noqa: F811
):
mock_brain = mb.setup_mock_brain(
use_discrete,
False,
vector_action_space=VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
discrete_action_space=DISCRETE_ACTION_SPACE,
)
trainer_params = dummy_config
trainer_params.network_settings.memory = NetworkSettings.MemorySettings(
memory_size=10, sequence_length=16
)
# Test curiosity reward signal
trainer_params.reward_signals = curiosity_dummy_config
trainer = PPOTrainer(mock_brain.brain_name, 0, trainer_params, True, False, 0, "0")
policy = trainer.create_policy(mock_brain.brain_name, mock_brain)
trainer.add_policy(mock_brain.brain_name, policy)
# Test update with sequence length smaller than batch size
buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, mock_brain)
# Mock out reward signal eval
buffer["extrinsic_rewards"] = buffer["environment_rewards"]
buffer["extrinsic_returns"] = buffer["environment_rewards"]
buffer["extrinsic_value_estimates"] = buffer["environment_rewards"]
buffer["curiosity_rewards"] = buffer["environment_rewards"]
buffer["curiosity_returns"] = buffer["environment_rewards"]
buffer["curiosity_value_estimates"] = buffer["environment_rewards"]
buffer["advantages"] = buffer["environment_rewards"]
trainer.update_buffer = buffer
trainer._update_policy()
def __init__(self, specs: BehaviorSpec,
settings: CuriositySettings) -> None:
super().__init__()
self._action_spec = specs.action_spec
state_encoder_settings = NetworkSettings(
normalize=False,
hidden_units=settings.encoding_size,
num_layers=2,
vis_encode_type=EncoderType.SIMPLE,
memory=None,
)
self._state_encoder = NetworkBody(specs.observation_specs,
state_encoder_settings)
self._action_flattener = ActionFlattener(self._action_spec)
self.inverse_model_action_encoding = torch.nn.Sequential(
LinearEncoder(2 * settings.encoding_size, 1, 256))
if self._action_spec.continuous_size > 0:
self.continuous_action_prediction = linear_layer(
256, self._action_spec.continuous_size)
if self._action_spec.discrete_size > 0:
self.discrete_action_prediction = linear_layer(
256, sum(self._action_spec.discrete_branches))
self.forward_model_next_state_prediction = torch.nn.Sequential(
LinearEncoder(
settings.encoding_size + self._action_flattener.flattened_size,
1, 256),
linear_layer(256, settings.encoding_size),
)
def test_normalizer_after_load(tmp_path):
behavior_spec = mb.setup_test_behavior_specs(
use_discrete=True, use_visual=False, vector_action_space=[2], vector_obs_space=1
)
time_horizon = 6
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
observation_shapes=[(1,)],
action_spec=behavior_spec.action_spec,
)
# Change half of the obs to 0
for i in range(3):
trajectory.steps[i].obs[0] = np.zeros(1, dtype=np.float32)
trainer_params = TrainerSettings(network_settings=NetworkSettings(normalize=True))
policy = TFPolicy(0, behavior_spec, trainer_params)
trajectory_buffer = trajectory.to_agentbuffer()
policy.update_normalization(trajectory_buffer["vector_obs"])
# Check that the running mean and variance is correct
steps, mean, variance = policy.sess.run(
[policy.normalization_steps, policy.running_mean, policy.running_variance]
)
assert steps == 6
assert mean[0] == 0.5
assert variance[0] / steps == pytest.approx(0.25, abs=0.01)
# Save ckpt and load into another policy
path1 = os.path.join(tmp_path, "runid1")
model_saver = TFModelSaver(trainer_params, path1)
model_saver.register(policy)
mock_brain_name = "MockBrain"
model_saver.save_checkpoint(mock_brain_name, 6)
assert len(os.listdir(tmp_path)) > 0
policy1 = TFPolicy(0, behavior_spec, trainer_params)
model_saver = TFModelSaver(trainer_params, path1, load=True)
model_saver.register(policy1)
model_saver.initialize_or_load(policy1)
# Make another update to new policy, this time with all 1's
time_horizon = 10
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
observation_shapes=[(1,)],
action_spec=behavior_spec.action_spec,
)
trajectory_buffer = trajectory.to_agentbuffer()
policy1.update_normalization(trajectory_buffer["vector_obs"])
# Check that the running mean and variance is correct
steps, mean, variance = policy1.sess.run(
[policy1.normalization_steps, policy1.running_mean, policy1.running_variance]
)
assert steps == 16
assert mean[0] == 0.8125
assert variance[0] / steps == pytest.approx(0.152, abs=0.01)
def create_optimizer_mock(trainer_config, reward_signal_config, use_rnn,
use_discrete, use_visual):
mock_specs = mb.setup_test_behavior_specs(
use_discrete,
use_visual,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE if not use_visual else 0,
)
trainer_settings = trainer_config
trainer_settings.reward_signals = reward_signal_config
trainer_settings.network_settings.memory = (NetworkSettings.MemorySettings(
sequence_length=16, memory_size=10) if use_rnn else None)
policy = NNPolicy(0,
mock_specs,
trainer_settings,
False,
"test",
False,
create_tf_graph=False)
if trainer_settings.trainer_type == TrainerType.SAC:
optimizer = SACOptimizer(policy, trainer_settings)
else:
optimizer = PPOOptimizer(policy, trainer_settings)
return optimizer
def test_multinetworkbody_num_agents(with_actions):
torch.manual_seed(0)
act_size = 2
obs_size = 4
network_settings = NetworkSettings()
obs_shapes = [(obs_size,)]
action_spec = ActionSpec(act_size, tuple(act_size for _ in range(act_size)))
networkbody = MultiAgentNetworkBody(
create_observation_specs_with_shapes(obs_shapes), network_settings, action_spec
)
sample_obs = [[0.1 * torch.ones((1, obs_size))]]
# simulate baseline in POCA
sample_act = [
AgentAction(
0.1 * torch.ones((1, 2)), [0.1 * torch.ones(1) for _ in range(act_size)]
)
]
for n_agent, max_so_far in [(1, 1), (5, 5), (4, 5), (10, 10), (5, 10), (1, 10)]:
if with_actions:
encoded, _ = networkbody(
obs_only=sample_obs * (n_agent - 1), obs=sample_obs, actions=sample_act
)
else:
encoded, _ = networkbody(obs_only=sample_obs * n_agent, obs=[], actions=[])
# look at the last value of the hidden units (the number of agents)
target = (n_agent * 1.0 / max_so_far) * 2 - 1
assert abs(encoded[0, -1].item() - target) < 1e-6
assert encoded[0, -1].item() <= 1
assert encoded[0, -1].item() >= -1
def test_recurrent_poca(action_sizes, is_multiagent):
if is_multiagent:
# This is not a recurrent environment, just check if LSTM doesn't crash
env = MultiAgentEnvironment([BRAIN_NAME],
action_sizes=action_sizes,
num_agents=2)
else:
# Actually test LSTM here
env = MemoryEnvironment([BRAIN_NAME], action_sizes=action_sizes)
new_network_settings = attr.evolve(
POCA_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16),
)
new_hyperparams = attr.evolve(
POCA_TORCH_CONFIG.hyperparameters,
learning_rate=1.0e-3,
batch_size=64,
buffer_size=128,
)
config = attr.evolve(
POCA_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=500 if is_multiagent else 6000,
)
check_environment_trains(env, {BRAIN_NAME: config},
success_threshold=None if is_multiagent else 0.9)
def create_policy_mock(
dummy_config: TrainerSettings,
use_rnn: bool = False,
use_discrete: bool = True,
use_visual: bool = False,
model_path: str = "",
load: bool = False,
seed: int = 0,
) -> TFPolicy:
mock_spec = mb.setup_test_behavior_specs(
use_discrete,
use_visual,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete
else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
)
trainer_settings = dummy_config
trainer_settings.keep_checkpoints = 3
trainer_settings.network_settings.memory = (
NetworkSettings.MemorySettings() if use_rnn else None
)
policy = TFPolicy(
seed, mock_spec, trainer_settings, model_path=model_path, load=load
)
return policy
def test_valuenetwork():
torch.manual_seed(0)
obs_size = 4
num_outputs = 2
network_settings = NetworkSettings()
obs_spec = create_observation_specs_with_shapes([(obs_size, )])
stream_names = [f"stream_name{n}" for n in range(4)]
value_net = ValueNetwork(stream_names,
obs_spec,
network_settings,
outputs_per_stream=num_outputs)
optimizer = torch.optim.Adam(value_net.parameters(), lr=3e-3)
for _ in range(50):
sample_obs = torch.ones((1, obs_size))
values, _ = value_net([sample_obs])
loss = 0
for s_name in stream_names:
assert values[s_name].shape == (1, num_outputs)
# Try to force output to 1
loss += torch.nn.functional.mse_loss(values[s_name],
torch.ones((1, num_outputs)))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# In the last step, values should be close to 1
for value in values.values():
for _out in value.tolist():
assert _out[0] == pytest.approx(1.0, abs=0.1)
def test_recurrent_sac(use_discrete):
step_size = 0.5 if use_discrete else 0.2
env = MemoryEnvironment([BRAIN_NAME],
use_discrete=use_discrete,
step_size=step_size)
new_networksettings = attr.evolve(
SAC_TF_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16,
sequence_length=16),
)
new_hyperparams = attr.evolve(
SAC_TF_CONFIG.hyperparameters,
batch_size=128,
learning_rate=1e-3,
buffer_init_steps=1000,
steps_per_update=2,
)
config = attr.evolve(
SAC_TF_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=5000,
framework=FrameworkType.TENSORFLOW,
)
_check_environment_trains(env, {BRAIN_NAME: config})
def __init__(self, specs: BehaviorSpec,
settings: CuriositySettings) -> None:
super().__init__()
self._policy_specs = specs
state_encoder_settings = NetworkSettings(
normalize=False,
hidden_units=settings.encoding_size,
num_layers=2,
vis_encode_type=EncoderType.SIMPLE,
memory=None,
)
self._state_encoder = NetworkBody(specs.observation_shapes,
state_encoder_settings)
self._action_flattener = ModelUtils.ActionFlattener(specs)
self.inverse_model_action_prediction = torch.nn.Sequential(
LinearEncoder(2 * settings.encoding_size, 1, 256),
linear_layer(256, self._action_flattener.flattened_size),
)
self.forward_model_next_state_prediction = torch.nn.Sequential(
LinearEncoder(
settings.encoding_size + self._action_flattener.flattened_size,
1, 256),
linear_layer(256, settings.encoding_size),
)
def test_networkbody_vector():
torch.manual_seed(0)
obs_size = 4
network_settings = NetworkSettings()
obs_shapes = [(obs_size, )]
networkbody = NetworkBody(
create_observation_specs_with_shapes(obs_shapes),
network_settings,
encoded_act_size=2,
)
optimizer = torch.optim.Adam(networkbody.parameters(), lr=3e-3)
sample_obs = 0.1 * torch.ones((1, obs_size))
sample_act = 0.1 * torch.ones((1, 2))
for _ in range(300):
encoded, _ = networkbody([sample_obs], sample_act)
assert encoded.shape == (1, network_settings.hidden_units)
# Try to force output to 1
loss = torch.nn.functional.mse_loss(encoded, torch.ones(encoded.shape))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# In the last step, values should be close to 1
for _enc in encoded.flatten():
assert _enc == pytest.approx(1.0, abs=0.1)
def test_actor_critic(ac_type, lstm):
obs_size = 4
network_settings = NetworkSettings(
memory=NetworkSettings.MemorySettings() if lstm else None)
obs_shapes = [(obs_size, )]
act_size = [2]
stream_names = [f"stream_name{n}" for n in range(4)]
actor = ac_type(obs_shapes, network_settings, ActionType.CONTINUOUS,
act_size, stream_names)
if lstm:
sample_obs = torch.ones(
(1, network_settings.memory.sequence_length, obs_size))
memories = torch.ones((
1,
network_settings.memory.sequence_length,
network_settings.memory.memory_size,
))
else:
sample_obs = torch.ones((1, obs_size))
memories = torch.tensor([])
# memories isn't always set to None, the network should be able to
# deal with that.
# Test critic pass
value_out, memories_out = actor.critic_pass([sample_obs], [],
memories=memories)
for stream in stream_names:
if lstm:
assert value_out[stream].shape == (
network_settings.memory.sequence_length, )
assert memories_out.shape == memories.shape
else:
assert value_out[stream].shape == (1, )
# Test get_dist_and_value
dists, value_out, mem_out = actor.get_dist_and_value([sample_obs], [],
memories=memories)
if mem_out is not None:
assert mem_out.shape == memories.shape
for dist in dists:
assert isinstance(dist, GaussianDistInstance)
for stream in stream_names:
if lstm:
assert value_out[stream].shape == (
network_settings.memory.sequence_length, )
else:
assert value_out[stream].shape == (1, )
def test_load_policy_different_hidden_units(tmp_path, vis_encode_type):
path1 = os.path.join(tmp_path, "runid1")
trainer_params = TrainerSettings()
trainer_params.network_settings = NetworkSettings(
hidden_units=12, vis_encode_type=EncoderType(vis_encode_type))
policy = create_policy_mock(trainer_params, use_visual=True)
conv_params = [
mod for mod in policy.actor.parameters() if len(mod.shape) > 2
]
model_saver = TorchModelSaver(trainer_params, path1)
model_saver.register(policy)
model_saver.initialize_or_load(policy)
policy.set_step(2000)
mock_brain_name = "MockBrain"
model_saver.save_checkpoint(mock_brain_name, 2000)
# Try load from this path
trainer_params2 = TrainerSettings()
trainer_params2.network_settings = NetworkSettings(
hidden_units=10, vis_encode_type=EncoderType(vis_encode_type))
model_saver2 = TorchModelSaver(trainer_params2, path1, load=True)
policy2 = create_policy_mock(trainer_params2, use_visual=True)
conv_params2 = [
mod for mod in policy2.actor.parameters() if len(mod.shape) > 2
]
# asserts convolutions have different parameters before load
for conv1, conv2 in zip(conv_params, conv_params2):
assert not torch.equal(conv1, conv2)
# asserts layers still have different dimensions
for mod1, mod2 in zip(policy.actor.parameters(),
policy2.actor.parameters()):
if mod1.shape[0] == 12:
assert mod2.shape[0] == 10
model_saver2.register(policy2)
model_saver2.initialize_or_load(policy2)
# asserts convolutions have same parameters after load
for conv1, conv2 in zip(conv_params, conv_params2):
assert torch.equal(conv1, conv2)
# asserts layers still have different dimensions
for mod1, mod2 in zip(policy.actor.parameters(),
policy2.actor.parameters()):
if mod1.shape[0] == 12:
assert mod2.shape[0] == 10
def __init__(self, specs: BehaviorSpec, settings: RNDSettings) -> None:
super().__init__()
state_encoder_settings = NetworkSettings(
normalize=True,
hidden_units=settings.encoding_size,
num_layers=3,
vis_encode_type=EncoderType.SIMPLE,
memory=None,
)
self._encoder = NetworkBody(specs.observation_specs, state_encoder_settings)
def test_trainer_update_policy(
dummy_config,
curiosity_dummy_config,
use_discrete # noqa: F811
):
mock_behavior_spec = mb.setup_test_behavior_specs(
use_discrete,
False,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
)
trainer_params = dummy_config
trainer_params.network_settings.memory = NetworkSettings.MemorySettings(
memory_size=10, sequence_length=16)
# Test curiosity reward signal
trainer_params.reward_signals = curiosity_dummy_config
mock_brain_name = "MockBrain"
behavior_id = BehaviorIdentifiers.from_name_behavior_id(mock_brain_name)
trainer = PPOTrainer("test", 0, trainer_params, True, False, 0, "0")
policy = trainer.create_policy(behavior_id, mock_behavior_spec)
trainer.add_policy(behavior_id, policy)
# Test update with sequence length smaller than batch size
buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, mock_behavior_spec)
# Mock out reward signal eval
buffer["extrinsic_rewards"] = buffer["environment_rewards"]
buffer["extrinsic_returns"] = buffer["environment_rewards"]
buffer["extrinsic_value_estimates"] = buffer["environment_rewards"]
buffer["curiosity_rewards"] = buffer["environment_rewards"]
buffer["curiosity_returns"] = buffer["environment_rewards"]
buffer["curiosity_value_estimates"] = buffer["environment_rewards"]
buffer["advantages"] = buffer["environment_rewards"]
# NOTE: This is because TF outputs the log probs of all actions whereas PyTorch does not
if use_discrete:
n_agents = len(buffer["discrete_log_probs"])
buffer["discrete_log_probs"].reset_field()
for _ in range(n_agents):
buffer["discrete_log_probs"].append(
np.ones(
int(sum(mock_behavior_spec.action_spec.discrete_branches)),
dtype=np.float32,
))
else:
n_agents = len(buffer["continuous_log_probs"])
buffer["continuous_log_probs"].reset_field()
for _ in range(n_agents):
buffer["continuous_log_probs"].append(
np.ones(mock_behavior_spec.action_spec.continuous_size,
dtype=np.float32))
trainer.update_buffer = buffer
trainer._update_policy()
def create_sac_optimizer_mock(dummy_config, use_rnn, use_discrete, use_visual):
mock_brain = mb.setup_test_behavior_specs(
use_discrete,
use_visual,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE if not use_visual else 0,
)
trainer_settings = dummy_config
trainer_settings.network_settings.memory = (NetworkSettings.MemorySettings(
sequence_length=16, memory_size=12) if use_rnn else None)
policy = TorchPolicy(0, mock_brain, trainer_settings)
optimizer = TorchSACOptimizer(policy, trainer_settings)
return optimizer
def convert_behaviors(old_trainer_config: Dict[str, Any]) -> Dict[str, Any]:
all_behavior_config_dict = {}
default_config = old_trainer_config.get("default", {})
for behavior_name, config in old_trainer_config.items():
if behavior_name != "default":
config = default_config.copy()
config.update(old_trainer_config[behavior_name])
# Convert to split TrainerSettings, Hyperparameters, NetworkSettings
# Set trainer_type and get appropriate hyperparameter settings
try:
trainer_type = config["trainer"]
except KeyError:
raise TrainerConfigError(
"Config doesn't specify a trainer type. "
"Please specify trainer: in your config."
)
new_config = {}
new_config["trainer_type"] = trainer_type
hyperparam_cls = TrainerType(trainer_type).to_settings()
# Try to absorb as much as possible into the hyperparam_cls
new_config["hyperparameters"] = cattr.structure(config, hyperparam_cls)
# Try to absorb as much as possible into the network settings
new_config["network_settings"] = cattr.structure(config, NetworkSettings)
# Deal with recurrent
try:
if config["use_recurrent"]:
new_config[
"network_settings"
].memory = NetworkSettings.MemorySettings(
sequence_length=config["sequence_length"],
memory_size=config["memory_size"],
)
except KeyError:
raise TrainerConfigError(
"Config doesn't specify use_recurrent. "
"Please specify true or false for use_recurrent in your config."
)
# Absorb the rest into the base TrainerSettings
for key, val in config.items():
if key in attr.fields_dict(TrainerSettings):
new_config[key] = val
# Structure the whole thing
all_behavior_config_dict[behavior_name] = cattr.structure(
new_config, TrainerSettings
)
return all_behavior_config_dict
def create_bc_module(mock_behavior_specs, bc_settings, use_rnn, tanhresample):
# model_path = env.external_brain_names[0]
trainer_config = TrainerSettings()
trainer_config.network_settings.memory = (NetworkSettings.MemorySettings()
if use_rnn else None)
policy = TorchPolicy(0, mock_behavior_specs, trainer_config, tanhresample,
tanhresample)
bc_module = BCModule(
policy,
settings=bc_settings,
policy_learning_rate=trainer_config.hyperparameters.learning_rate,
default_batch_size=trainer_config.hyperparameters.batch_size,
default_num_epoch=3,
)
return bc_module
def create_test_ppo_optimizer(dummy_config, use_rnn, use_discrete, use_visual):
mock_specs = mb.setup_test_behavior_specs(
use_discrete,
use_visual,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
)
trainer_settings = attr.evolve(dummy_config)
trainer_settings.network_settings.memory = (NetworkSettings.MemorySettings(
sequence_length=16, memory_size=10) if use_rnn else None)
policy = TorchPolicy(0, mock_specs, trainer_settings, "test", False)
optimizer = TorchPPOOptimizer(policy, trainer_settings)
return optimizer
def __init__(self, specs: BehaviorSpec, settings: GAILSettings) -> None:
super().__init__()
self._policy_specs = specs
self._use_vail = settings.use_vail
self._settings = settings
state_encoder_settings = NetworkSettings(
normalize=False,
hidden_units=settings.encoding_size,
num_layers=2,
vis_encode_type=EncoderType.SIMPLE,
memory=None,
)
self._state_encoder = NetworkBody(specs.observation_shapes,
state_encoder_settings)
self._action_flattener = ModelUtils.ActionFlattener(specs)
encoder_input_size = settings.encoding_size
if settings.use_actions:
encoder_input_size += (self._action_flattener.flattened_size + 1
) # + 1 is for done
self.encoder = torch.nn.Sequential(
linear_layer(encoder_input_size, settings.encoding_size),
Swish(),
linear_layer(settings.encoding_size, settings.encoding_size),
Swish(),
)
estimator_input_size = settings.encoding_size
if settings.use_vail:
estimator_input_size = self.z_size
self._z_sigma = torch.nn.Parameter(torch.ones((self.z_size),
dtype=torch.float),
requires_grad=True)
self._z_mu_layer = linear_layer(
settings.encoding_size,
self.z_size,
kernel_init=Initialization.KaimingHeNormal,
kernel_gain=0.1,
)
self._beta = torch.nn.Parameter(torch.tensor(self.initial_beta,
dtype=torch.float),
requires_grad=False)
self._estimator = torch.nn.Sequential(
linear_layer(estimator_input_size, 1), torch.nn.Sigmoid())
def test_recurrent_ppo(use_discrete):
env = MemoryEnvironment([BRAIN_NAME], use_discrete=use_discrete)
new_network_settings = attr.evolve(
PPO_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16),
)
new_hyperparams = attr.evolve(
PPO_CONFIG.hyperparameters, learning_rate=1.0e-3, batch_size=64, buffer_size=128
)
config = attr.evolve(
PPO_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=5000,
)
_check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.9)
def test_hybrid_recurrent_ppo():
env = MemoryEnvironment([BRAIN_NAME], action_sizes=(1, 1), step_size=0.5)
new_network_settings = attr.evolve(
PPO_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16),
)
new_hyperparams = attr.evolve(
PPO_TORCH_CONFIG.hyperparameters,
learning_rate=1.0e-3,
batch_size=64,
buffer_size=512,
)
config = attr.evolve(
PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=3000,
)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.9)
def create_sac_optimizer_mock(dummy_config, use_rnn, use_discrete, use_visual):
mock_brain = mb.setup_mock_brain(
use_discrete,
use_visual,
vector_action_space=VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
discrete_action_space=DISCRETE_ACTION_SPACE,
)
trainer_settings = dummy_config
trainer_settings.network_settings.memory = (
NetworkSettings.MemorySettings(sequence_length=16, memory_size=10)
if use_rnn
else None
)
policy = NNPolicy(
0, mock_brain, trainer_settings, False, "test", False, create_tf_graph=False
)
optimizer = SACOptimizer(policy, trainer_settings)
return optimizer
