def evaluate_gym(
env_name: str,
model: ModelManager__Union,
publisher: ModelPublisher__Union,
num_eval_episodes: int,
passing_score_bar: float,
max_steps: Optional[int] = None,
):
publisher_manager = publisher.value
assert isinstance(
publisher_manager, FileSystemPublisher
), f"publishing manager is type {type(publisher_manager)}, not FileSystemPublisher"
env = EnvFactory.make(env_name)
torchscript_path = publisher_manager.get_latest_published_model(
model.value)
jit_model = torch.jit.load(torchscript_path)
policy = create_predictor_policy_from_model(jit_model)
agent = Agent.create_for_env_with_serving_policy(env, policy)
rewards = evaluate_for_n_episodes(n=num_eval_episodes,
env=env,
agent=agent,
max_steps=max_steps)
avg_reward = np.mean(rewards)
logger.info(f"Average reward over {num_eval_episodes} is {avg_reward}.\n"
f"List of rewards: {rewards}")
assert (avg_reward >= passing_score_bar
), f"{avg_reward} fails to pass the bar of {passing_score_bar}!"
return
def test_random_vs_lqr(self):
"""
Test random actions vs. a LQR controller. LQR controller should perform
much better than random actions in the linear dynamics environment.
"""
env = EnvFactory.make("LinearDynamics-v0")
num_test_episodes = 500
def random_policy(env, state):
return np.random.uniform(env.action_space.low,
env.action_space.high, env.action_dim)
def lqr_policy(env, state):
# Four matrices that characterize the environment
A, B, Q, R = env.A, env.B, env.Q, env.R
# Solve discrete algebraic Riccati equation:
M = linalg.solve_discrete_are(A, B, Q, R)
K = np.dot(linalg.inv(np.dot(np.dot(B.T, M), B) + R),
(np.dot(np.dot(B.T, M), A)))
state = state.reshape((-1, 1))
action = -K.dot(state).squeeze()
return action
mean_acc_rws_random = self.run_n_episodes(env, num_test_episodes,
random_policy)
mean_acc_rws_lqr = self.run_n_episodes(env, num_test_episodes,
lqr_policy)
logger.info(
f"Mean acc. reward of random policy: {mean_acc_rws_random}")
logger.info(f"Mean acc. reward of LQR policy: {mean_acc_rws_lqr}")
assert mean_acc_rws_lqr > mean_acc_rws_random
def offline_gym(
env: str,
pkl_path: str,
num_episodes_for_data_batch: int,
max_steps: Optional[int],
seed: Optional[int] = None,
):
"""
Generate samples from a DiscreteRandomPolicy on the Gym environment and
saves results in a pandas df parquet.
"""
initialize_seed(seed)
env = EnvFactory.make(env)
policy = DiscreteRandomPolicy.create_for_env(env)
dataset = RLDataset()
for i in range(num_episodes_for_data_batch):
logger.info(f"Starting episode {i}")
post_step = log_data_post_step(dataset=dataset, mdp_id=str(i), env=env)
agent = Agent.create_for_env(env,
policy,
post_transition_callback=post_step)
run_episode(env=env, agent=agent, max_steps=max_steps)
logger.info(f"Saving dataset with {len(dataset)} samples to {pkl_path}")
df = dataset.to_pandas_df()
df.to_pickle(pkl_path)
def evaluate_gym(
env_name: str,
model: torch.nn.Module,
eval_temperature: float,
num_eval_episodes: int,
passing_score_bar: float,
max_steps: Optional[int] = None,
):
env: gym.Env = EnvFactory.make(env_name)
policy = create_predictor_policy_from_model(
env, model, eval_temperature=eval_temperature)
# since we already return softmax action, override action_extractor
agent = Agent.create_for_env(
env, policy=policy, action_extractor=policy.get_action_extractor())
rewards = []
for _ in range(num_eval_episodes):
ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps)
rewards.append(ep_reward)
avg_reward = np.mean(rewards)
logger.info(f"Average reward over {num_eval_episodes} is {avg_reward}.\n"
f"List of rewards: {rewards}")
assert (avg_reward >= passing_score_bar
), f"{avg_reward} fails to pass the bar of {passing_score_bar}!"
def evaluate_gym(
env: str,
model,
eval_temperature: float,
num_eval_episodes: int,
passing_score_bar: float,
max_steps: Optional[int] = None,
):
predictor = DiscreteDqnTorchPredictor(model)
predictor.softmax_temperature = eval_temperature
env = EnvFactory.make(env)
policy = TorchPredictorPolicy(predictor)
agent = Agent(policy=policy, action_extractor=lambda x: x.item())
rewards = []
for _ in range(num_eval_episodes):
ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps)
rewards.append(ep_reward)
avg_reward = np.mean(rewards)
logger.info(f"Average reward over {num_eval_episodes} is {avg_reward}, "
f"which passes the bar of {passing_score_bar}!\n"
f"List of rewards: {rewards}")
assert avg_reward >= passing_score_bar
def offline_gym(
env_name: str,
pkl_path: str,
num_train_transitions: int,
max_steps: Optional[int],
seed: Optional[int] = None,
):
"""
Generate samples from a DiscreteRandomPolicy on the Gym environment and
saves results in a pandas df parquet.
"""
initialize_seed(seed)
env = EnvFactory.make(env_name)
replay_buffer = ReplayBuffer.create_from_env(
env=env, replay_memory_size=num_train_transitions, batch_size=1)
fill_replay_buffer(env, replay_buffer, num_train_transitions)
if isinstance(env.action_space, gym.spaces.Discrete):
is_discrete_action = True
else:
assert isinstance(env.action_space, gym.spaces.Box)
is_discrete_action = False
df = replay_buffer_to_pre_timeline_df(is_discrete_action, replay_buffer)
logger.info(f"Saving dataset with {len(df)} samples to {pkl_path}")
df.to_pickle(pkl_path)
def run_test_offline(
env_name: str,
max_steps: Optional[int],
model: ModelManager__Union,
replay_memory_size: int,
num_batches_per_epoch: int,
num_train_epochs: int,
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
env = EnvFactory.make(env_name)
env.seed(SEED)
env.action_space.seed(SEED)
normalization = build_normalizer(env)
logger.info(f"Normalization is: \n{pprint.pformat(normalization)}")
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=normalization,
)
# first fill the replay buffer to burn_in
replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=replay_memory_size,
batch_size=trainer.minibatch_size,
)
# always fill full RB
fill_replay_buffer(
env=env, replay_buffer=replay_buffer, desired_size=replay_memory_size
)
device = torch.device("cuda") if use_gpu else None
# pyre-fixme[6]: Expected `device` for 2nd param but got `Optional[torch.device]`.
trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env)
writer = SummaryWriter()
with summary_writer_context(writer):
for epoch in range(num_train_epochs):
logger.info(f"Evaluating before epoch {epoch}: ")
eval_rewards = evaluate_cem(env, manager, max_steps, 1)
for _ in tqdm(range(num_batches_per_epoch)):
train_batch = replay_buffer.sample_transition_batch_tensor()
preprocessed_batch = trainer_preprocessor(train_batch)
trainer.train(preprocessed_batch)
logger.info(f"Evaluating after training for {num_train_epochs} epochs: ")
eval_rewards = evaluate_cem(env, manager, max_steps, num_eval_episodes)
mean_rewards = np.mean(eval_rewards)
assert (
mean_rewards >= passing_score_bar
), f"{mean_rewards} doesn't pass the bar {passing_score_bar}."
def run_test(
env: str,
model: ModelManager__Union,
replay_memory_size: int,
train_every_ts: int,
train_after_ts: int,
num_episodes: int,
max_steps: Optional[int],
last_score_bar: float,
):
env = EnvFactory.make(env)
env.seed(SEED)
env.action_space.seed(SEED)
normalization = build_normalizer(env)
logger.info(f"Normalization is {normalization}")
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=False,
reward_options=RewardOptions(),
normalization_data_map=normalization,
)
policy = manager.create_policy()
replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=replay_memory_size,
batch_size=trainer.minibatch_size,
)
post_step = train_with_replay_buffer_post_step(
replay_buffer=replay_buffer,
trainer=trainer,
training_freq=train_every_ts,
batch_size=trainer.minibatch_size,
replay_burnin=train_after_ts,
)
agent = Agent.create_for_env(env,
policy=policy,
post_transition_callback=post_step)
reward_history = []
for i in range(num_episodes):
logger.info(f"running episode {i}")
ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps)
reward_history.append(ep_reward)
assert reward_history[-1] >= last_score_bar, (
f"reward after {len(reward_history)} episodes is {reward_history[-1]},"
f"less than < {last_score_bar}...\n"
f"Full reward history: {reward_history}")
return reward_history
def _create_env(self, gymenv: Union[str, Env], random_seed: Optional[int]):
"""
Creates a gym environment object and checks if it is supported. We
support environments that supply Box(x, ) state representations and
require Discrete(y) or Box(y,) action inputs.
:param gymenv: String identifier for desired environment or environment
object itself.
"""
if isinstance(gymenv, Env):
self.env = gymenv
self.env_name = gymenv.unwrapped.spec.id
else:
if gymenv not in [e.id for e in gym.envs.registry.all()]:
raise Exception(
"Env {} not found in OpenAI Gym.".format(gymenv))
self.env = EnvFactory.make(gymenv)
self.env_name = gymenv
if random_seed is not None:
self.env.seed(random_seed)
self.env.action_space.seed(random_seed) # type: ignore
supports_state = isinstance(
self.env.observation_space, gym.spaces.Box) and len(
self.env.observation_space.shape) in [1, 3]
supports_action = type(self.env.action_space) in (
gym.spaces.Discrete,
gym.spaces.Box,
)
if not supports_state and supports_action:
raise Exception(
"Unsupported environment state or action type: {}, {}".format(
self.env.observation_space, self.env.action_space))
self.action_space = self.env.action_space
if isinstance(self.env.action_space, gym.spaces.Discrete):
self.action_type = EnvType.DISCRETE_ACTION
self.action_dim = self.env.action_space.n
elif isinstance(self.env.action_space, gym.spaces.Box):
self.action_type = EnvType.CONTINUOUS_ACTION
self.action_dim = self.env.action_space.shape[0] # type: ignore
if len(self.env.observation_space.shape) == 1: # type: ignore
self.state_dim = self.env.observation_space.shape[
0] # type: ignore
self.img = False
elif len(self.env.observation_space.shape) == 3: # type: ignore
self.height, self.width, self.num_input_channels = (
self.env.observation_space.shape # type: ignore
)
self.img = True
def train_mdnrnn_and_train_on_embedded_env(
env_name: str,
embedding_model: ModelManager__Union,
num_embedding_train_transitions: int,
seq_len: int,
batch_size: int,
num_embedding_train_epochs: int,
train_model: ModelManager__Union,
num_state_embed_transitions: int,
num_agent_train_epochs: int,
num_agent_eval_epochs: int,
use_gpu: bool,
passing_score_bar: float,
# pyre-fixme[9]: saved_mdnrnn_path has type `str`; used as `None`.
saved_mdnrnn_path: str = None,
):
""" Train an agent on embedded states by the MDNRNN. """
env = EnvFactory.make(env_name)
env.seed(SEED)
embedding_manager = embedding_model.value
embedding_trainer = embedding_manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(env),
)
device = "cuda" if use_gpu else "cpu"
embedding_trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
embedding_trainer, device, env
)
if saved_mdnrnn_path is None:
# train from scratch
embedding_trainer = train_mdnrnn(
env=env,
trainer=embedding_trainer,
trainer_preprocessor=embedding_trainer_preprocessor,
num_train_transitions=num_embedding_train_transitions,
seq_len=seq_len,
batch_size=batch_size,
num_train_epochs=num_embedding_train_epochs,
)
else:
# load a pretrained model, and just evaluate it
embedding_trainer.memory_network.mdnrnn.load_state_dict(
torch.load(saved_mdnrnn_path)
)
# create embedding dataset
embed_rb, state_min, state_max = create_embed_rl_dataset(
env=env,
memory_network=embedding_trainer.memory_network,
num_state_embed_transitions=num_state_embed_transitions,
batch_size=batch_size,
seq_len=seq_len,
hidden_dim=embedding_trainer.params.hidden_size,
use_gpu=use_gpu,
)
embed_env = StateEmbedEnvironment(
gym_env=env,
mdnrnn=embedding_trainer.memory_network,
max_embed_seq_len=seq_len,
state_min_value=state_min,
state_max_value=state_max,
)
agent_manager = train_model.value
agent_trainer = agent_manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(embed_env),
)
device = "cuda" if use_gpu else "cpu"
agent_trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
agent_trainer, device, env
)
num_batch_per_epoch = embed_rb.size // batch_size
for epoch in range(num_agent_train_epochs):
for _ in tqdm(range(num_batch_per_epoch), desc=f"epoch {epoch}"):
batch = embed_rb.sample_transition_batch_tensor(batch_size=batch_size)
preprocessed_batch = agent_trainer_preprocessor(batch)
agent_trainer.train(preprocessed_batch)
# evaluate model
rewards = []
policy = agent_manager.create_policy(serving=False)
agent = Agent.create_for_env(embed_env, policy=policy, device=device)
for i in range(num_agent_eval_epochs):
ep_reward = run_episode(env=embed_env, agent=agent)
rewards.append(ep_reward)
logger.info(f"Finished eval episode {i} with reward {ep_reward}.")
logger.info(f"Average eval reward is {np.mean(rewards)}.")
assert (
np.mean(rewards) >= passing_score_bar
), f"average reward doesn't pass our bar {passing_score_bar}"
return rewards
def train_mdnrnn_and_compute_feature_stats(
env_name: str,
model: ModelManager__Union,
num_train_transitions: int,
num_test_transitions: int,
seq_len: int,
batch_size: int,
num_train_epochs: int,
use_gpu: bool,
saved_mdnrnn_path: Optional[str] = None,
):
""" Train MDNRNN Memory Network and compute feature importance/sensitivity. """
env: gym.Env = EnvFactory.make(env_name)
env.seed(SEED)
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(env),
)
device = "cuda" if use_gpu else "cpu"
trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env)
test_replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=num_test_transitions,
batch_size=batch_size,
stack_size=seq_len,
return_everything_as_stack=True,
)
fill_replay_buffer(env, test_replay_buffer, num_test_transitions)
if saved_mdnrnn_path is None:
# train from scratch
trainer = train_mdnrnn(
env=env,
trainer=trainer,
trainer_preprocessor=trainer_preprocessor,
num_train_transitions=num_train_transitions,
seq_len=seq_len,
batch_size=batch_size,
num_train_epochs=num_train_epochs,
test_replay_buffer=test_replay_buffer,
)
else:
# load a pretrained model, and just evaluate it
trainer.memory_network.mdnrnn.load_state_dict(torch.load(saved_mdnrnn_path))
with torch.no_grad():
trainer.memory_network.mdnrnn.eval()
test_batch = test_replay_buffer.sample_transition_batch_tensor(
batch_size=test_replay_buffer.size
)
preprocessed_test_batch = trainer_preprocessor(test_batch)
feature_importance = calculate_feature_importance(
env=env,
trainer=trainer,
use_gpu=use_gpu,
test_batch=preprocessed_test_batch,
)
feature_sensitivity = calculate_feature_sensitivity(
env=env,
trainer=trainer,
use_gpu=use_gpu,
test_batch=preprocessed_test_batch,
)
trainer.memory_network.mdnrnn.train()
return feature_importance, feature_sensitivity
def run_test(
env: str,
model: ModelManager__Union,
replay_memory_size: int,
train_every_ts: int,
train_after_ts: int,
num_train_episodes: int,
max_steps: Optional[int],
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
env = EnvFactory.make(env)
env.seed(SEED)
env.action_space.seed(SEED)
normalization = build_normalizer(env)
logger.info(f"Normalization is {normalization}")
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=normalization,
)
replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=replay_memory_size,
batch_size=trainer.minibatch_size,
)
device = torch.device("cuda") if use_gpu else None
post_step = train_with_replay_buffer_post_step(
replay_buffer=replay_buffer,
trainer=trainer,
training_freq=train_every_ts,
batch_size=trainer.minibatch_size,
replay_burnin=train_after_ts,
device=device,
)
training_policy = manager.create_policy(serving=False)
agent = Agent.create_for_env(env,
policy=training_policy,
post_transition_callback=post_step,
device=device)
train_rewards = []
for i in range(num_train_episodes):
ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps)
train_rewards.append(ep_reward)
logger.info(f"Finished training episode {i} with reward {ep_reward}.")
assert train_rewards[-1] >= passing_score_bar, (
f"reward after {len(train_rewards)} episodes is {train_rewards[-1]},"
f"less than < {passing_score_bar}...\n"
f"Full reward history: {train_rewards}")
logger.info("============Train rewards=============")
logger.info(train_rewards)
def gym_to_reagent_serving(
obs: np.array) -> Tuple[torch.Tensor, torch.Tensor]:
obs_tensor = torch.tensor(obs).float().unsqueeze(0)
presence_tensor = torch.ones_like(obs_tensor)
return (obs_tensor, presence_tensor)
serving_policy = manager.create_policy(serving=True)
agent = Agent.create_for_env(env,
policy=serving_policy,
obs_preprocessor=gym_to_reagent_serving)
eval_rewards = []
for i in range(num_eval_episodes):
ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps)
eval_rewards.append(ep_reward)
logger.info(f"Finished eval episode {i} with reward {ep_reward}.")
assert np.mean(eval_rewards) >= passing_score_bar, (
f"Predictor reward is {np.mean(eval_rewards)},"
f"less than < {passing_score_bar}...\n"
f"Full eval rewards: {eval_rewards}.")
logger.info("============Eval rewards==============")
logger.info(eval_rewards)
def test_pocman(self):
env = EnvFactory.make("Pocman-v0")
env.seed(313)
mean_acc_reward = self._test_env(env)
assert -80 <= mean_acc_reward <= -70
def test_string_game(self):
env = EnvFactory.make("StringGame-v0")
env.seed(313)
mean_acc_reward = self._test_env(env)
assert 0.1 >= mean_acc_reward
def run_test(
env_name: str,
model: ModelManager__Union,
replay_memory_size: int,
train_every_ts: int,
train_after_ts: int,
num_train_episodes: int,
max_steps: Optional[int],
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
env = EnvFactory.make(env_name)
env.seed(SEED)
env.action_space.seed(SEED)
normalization = build_normalizer(env)
logger.info(f"Normalization is: \n{pprint.pformat(normalization)}")
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=normalization,
)
training_policy = manager.create_policy(serving=False)
replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=replay_memory_size,
batch_size=trainer.minibatch_size,
)
device = torch.device("cuda") if use_gpu else None
# first fill the replay buffer to burn_in
train_after_ts = max(train_after_ts, trainer.minibatch_size)
fill_replay_buffer(env=env,
replay_buffer=replay_buffer,
desired_size=train_after_ts)
post_step = train_with_replay_buffer_post_step(
replay_buffer=replay_buffer,
env=env,
trainer=trainer,
training_freq=train_every_ts,
batch_size=trainer.minibatch_size,
device=device,
)
agent = Agent.create_for_env(
env,
policy=training_policy,
post_transition_callback=post_step,
# pyre-fixme[6]: Expected `Union[str, torch.device]` for 4th param but got
# `Optional[torch.device]`.
device=device,
)
writer = SummaryWriter()
with summary_writer_context(writer):
train_rewards = []
for i in range(num_train_episodes):
trajectory = run_episode(env=env,
agent=agent,
mdp_id=i,
max_steps=max_steps)
ep_reward = trajectory.calculate_cumulative_reward()
train_rewards.append(ep_reward)
logger.info(
f"Finished training episode {i} with reward {ep_reward}.")
assert train_rewards[-1] >= passing_score_bar, (
f"reward after {len(train_rewards)} episodes is {train_rewards[-1]},"
f"less than < {passing_score_bar}...\n"
f"Full reward history: {train_rewards}")
logger.info("============Train rewards=============")
logger.info(train_rewards)
serving_policy = manager.create_policy(serving=True)
agent = Agent.create_for_env_with_serving_policy(env, serving_policy)
eval_rewards = evaluate_for_n_episodes(n=num_eval_episodes,
env=env,
agent=agent,
max_steps=max_steps).squeeze(1)
assert np.mean(eval_rewards) >= passing_score_bar, (
f"Predictor reward is {np.mean(eval_rewards)},"
f"less than < {passing_score_bar}...\n"
f"Full eval rewards: {eval_rewards}.")
logger.info("============Eval rewards==============")
logger.info(eval_rewards)
def run_test(
env_name: str,
model: ModelManager__Union,
replay_memory_size: int,
train_every_ts: int,
train_after_ts: int,
num_train_episodes: int,
max_steps: Optional[int],
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
env = EnvFactory.make(env_name)
env.seed(SEED)
env.action_space.seed(SEED)
normalization = build_normalizer(env)
logger.info(f"Normalization is: \n{pprint.pformat(normalization)}")
manager = model.value
try:
# pyre-fixme[16]: `Env` has no attribute `state_feature_config_provider`.
manager.state_feature_config_provider = env.state_feature_config_provider
logger.info(
f"Using environment's state_feature_config_provider.\n"
f"{manager.state_feature_config_provider}"
)
except AttributeError:
logger.info("state_feature_config_provider override not applicable")
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=normalization,
)
training_policy = manager.create_policy(serving=False)
replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=replay_memory_size,
batch_size=trainer.minibatch_size,
)
device = torch.device("cuda") if use_gpu else torch.device("cpu")
# first fill the replay buffer to burn_in
train_after_ts = max(train_after_ts, trainer.minibatch_size)
fill_replay_buffer(
env=env, replay_buffer=replay_buffer, desired_size=train_after_ts
)
post_step = train_with_replay_buffer_post_step(
replay_buffer=replay_buffer,
env=env,
trainer=trainer,
training_freq=train_every_ts,
batch_size=trainer.minibatch_size,
device=device,
)
agent = Agent.create_for_env(
env, policy=training_policy, post_transition_callback=post_step, device=device
)
writer = SummaryWriter()
with summary_writer_context(writer):
train_rewards = []
for i in range(num_train_episodes):
trajectory = run_episode(
env=env, agent=agent, mdp_id=i, max_steps=max_steps
)
ep_reward = trajectory.calculate_cumulative_reward()
train_rewards.append(ep_reward)
logger.info(
f"Finished training episode {i} (len {len(trajectory)})"
f" with reward {ep_reward}."
)
logger.info("============Train rewards=============")
logger.info(train_rewards)
logger.info(f"average: {np.mean(train_rewards)};\tmax: {np.max(train_rewards)}")
# Check whether the max score passed the score bar; we explore during training
# the return could be bad (leading to flakiness in C51 and QRDQN).
assert np.max(train_rewards) >= passing_score_bar, (
f"max reward ({np.max(train_rewards)})after training for "
f"{len(train_rewards)} episodes is less than < {passing_score_bar}.\n"
)
serving_policy = manager.create_policy(serving=True)
agent = Agent.create_for_env_with_serving_policy(env, serving_policy)
eval_rewards = evaluate_for_n_episodes(
n=num_eval_episodes, env=env, agent=agent, max_steps=max_steps
).squeeze(1)
logger.info("============Eval rewards==============")
logger.info(eval_rewards)
logger.info(f"average: {np.mean(eval_rewards)};\tmax: {np.max(eval_rewards)}")
assert np.mean(eval_rewards) >= passing_score_bar, (
f"Predictor reward is {np.mean(eval_rewards)},"
f"less than < {passing_score_bar}.\n"
)
def train_seq2reward_and_compute_reward_mse(
env_name: str,
model: ModelManager__Union,
num_train_transitions: int,
num_test_transitions: int,
seq_len: int,
batch_size: int,
num_train_epochs: int,
use_gpu: bool,
saved_seq2reward_path: Optional[str] = None,
):
""" Train Seq2Reward Network and compute reward mse. """
env: gym.Env = EnvFactory.make(env_name)
env.seed(SEED)
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(env),
)
device = "cuda" if use_gpu else "cpu"
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env)
test_replay_buffer = ReplayBuffer.create_from_env(
env=env,
replay_memory_size=num_test_transitions,
batch_size=batch_size,
stack_size=seq_len,
return_everything_as_stack=True,
)
fill_replay_buffer(env, test_replay_buffer, num_test_transitions)
if saved_seq2reward_path is None:
# train from scratch
trainer = train_seq2reward(
env=env,
trainer=trainer,
trainer_preprocessor=trainer_preprocessor,
num_train_transitions=num_train_transitions,
seq_len=seq_len,
batch_size=batch_size,
num_train_epochs=num_train_epochs,
test_replay_buffer=test_replay_buffer,
)
else:
# load a pretrained model, and just evaluate it
trainer.seq2reward_network.load_state_dict(torch.load(saved_seq2reward_path))
state_dim = env.observation_space.shape[0]
with torch.no_grad():
trainer.seq2reward_network.eval()
test_batch = test_replay_buffer.sample_transition_batch(
batch_size=test_replay_buffer.size
)
preprocessed_test_batch = trainer_preprocessor(test_batch)
adhoc_action_padding(preprocessed_test_batch, state_dim=state_dim)
losses = trainer.get_loss(preprocessed_test_batch)
detached_losses = losses.cpu().detach().item()
trainer.seq2reward_network.train()
return detached_losses
