def evaluate_cem(env, manager, num_eval_episodes: int):
# NOTE: for CEM, serving isn't implemented
policy = manager.create_policy(serving=False)
agent = Agent.create_for_env(env, policy)
return evaluate_for_n_episodes(
n=num_eval_episodes, env=env, agent=agent, max_steps=env.max_steps
)
def eval_policy(
env: EnvWrapper,
serving_policy: Policy,
num_eval_episodes: int,
serving: bool = True,
) -> np.ndarray:
agent = (
Agent.create_for_env_with_serving_policy(env, serving_policy)
if serving
else Agent.create_for_env(env, serving_policy)
)
eval_rewards = evaluate_for_n_episodes(
n=num_eval_episodes,
env=env,
agent=agent,
max_steps=env.max_steps,
num_processes=1,
).squeeze(1)
logger.info("============Eval rewards==============")
logger.info(eval_rewards)
mean_eval = np.mean(eval_rewards)
logger.info(f"average: {mean_eval};\tmax: {np.max(eval_rewards)}")
return np.array(eval_rewards)
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 = Gym(env_name=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 run_test_online_episode(
env: Env__Union,
model: ModelManager__Union,
num_train_episodes: int,
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
"""
Run an online learning test. At the end of each episode training is run on the trajectory.
"""
env = env.value
pl.seed_everything(SEED)
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.build_trainer(
use_gpu=use_gpu,
normalization_data_map=normalization,
)
policy = manager.create_policy(trainer, serving=False)
device = torch.device("cuda") if use_gpu else torch.device("cpu")
agent = Agent.create_for_env(env, policy, device=device)
pl_trainer = pl.Trainer(
max_epochs=1,
gpus=int(use_gpu),
deterministic=True,
default_root_dir=f"lightning_log_{str(uuid.uuid4())}",
)
dataset = EpisodicDataset(env=env,
agent=agent,
num_episodes=num_train_episodes,
seed=SEED)
data_loader = torch.utils.data.DataLoader(dataset,
collate_fn=identity_collate)
pl_trainer.fit(trainer, data_loader)
eval_rewards = evaluate_for_n_episodes(
n=num_eval_episodes,
env=env,
agent=agent,
max_steps=env.max_steps,
num_processes=1,
).squeeze(1)
assert (
eval_rewards.mean() >= passing_score_bar
), f"Eval reward is {eval_rewards.mean()}, less than < {passing_score_bar}.\n"
def evaluate_gym(
env_name: str,
model: ModelManager__Union,
publisher: ModelPublisher__Union,
num_eval_episodes: int,
passing_score_bar: float,
module_name: str = "default_model",
max_steps: Optional[int] = None,
):
initialize_seed(1)
env = Gym(env_name=env_name)
agent = make_agent_from_model(env, model, publisher, module_name)
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}\n"
f"Passing score bar: {passing_score_bar}")
assert (avg_reward >= passing_score_bar
), f"{avg_reward} fails to pass the bar of {passing_score_bar}!"
return
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 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 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 = Gym(env_name=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,
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
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(),
# pyre-fixme[6]: Expected `EnvWrapper` for 1st param but got
# `StateEmbedEnvironment`.
normalization_data_map=build_normalizer(embed_env),
)
device = "cuda" if use_gpu else "cpu"
agent_trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
agent_trainer,
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
device,
env,
)
num_batch_per_epoch = embed_rb.size // batch_size
# FIXME: This has to be wrapped in dataloader
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(batch_size=batch_size)
preprocessed_batch = agent_trainer_preprocessor(batch)
# FIXME: This should be fitted with Lightning's trainer
agent_trainer.train(preprocessed_batch)
# evaluate model
rewards = []
policy = agent_manager.create_policy(serving=False)
# pyre-fixme[6]: Expected `EnvWrapper` for 1st param but got
# `StateEmbedEnvironment`.
agent = Agent.create_for_env(embed_env, policy=policy, device=device)
# num_processes=1 needed to avoid workers from dying on CircleCI tests
rewards = evaluate_for_n_episodes(
n=num_agent_eval_epochs,
# pyre-fixme[6]: Expected `EnvWrapper` for 2nd param but got
# `StateEmbedEnvironment`.
env=embed_env,
agent=agent,
num_processes=1,
)
assert (np.mean(rewards) >= passing_score_bar
), f"average reward doesn't pass our bar {passing_score_bar}"
return rewards
def run_test_online_episode(
env: Env__Union,
model: ModelManager__Union,
num_train_episodes: int,
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
"""
Run an online learning test. At the end of each episode training is run on the trajectory.
"""
env = env.value
# pyre-fixme[16]: Module `pl` has no attribute `seed_everything`.
pl.seed_everything(SEED)
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,
)
policy = manager.create_policy(serving=False)
device = torch.device("cuda") if use_gpu else torch.device("cpu")
agent = Agent.create_for_env(env, policy, device=device)
# pyre-fixme[16]: Module `pl` has no attribute `LightningModule`.
if isinstance(trainer, pl.LightningModule):
# pyre-fixme[16]: Module `pl` has no attribute `Trainer`.
pl_trainer = pl.Trainer(max_epochs=1,
gpus=int(use_gpu),
deterministic=True)
dataset = EpisodicDataset(env=env,
agent=agent,
num_episodes=num_train_episodes,
seed=SEED)
pl_trainer.fit(trainer, dataset)
else:
post_episode_callback = train_post_episode(env, trainer, use_gpu)
_ = train_policy(
env,
policy,
num_train_episodes,
post_step=None,
post_episode=post_episode_callback,
use_gpu=use_gpu,
)
eval_rewards = evaluate_for_n_episodes(
n=num_eval_episodes,
env=env,
agent=agent,
max_steps=env.max_steps,
num_processes=1,
).squeeze(1)
assert (
eval_rewards.mean() >= passing_score_bar
), f"Eval reward is {eval_rewards.mean()}, less than < {passing_score_bar}.\n"
