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 fill_replay_buffer(env, replay_buffer: ReplayBuffer, desired_size: int):
""" Fill replay buffer with random transitions until size reaches desired_size. """
assert (
0 < desired_size and desired_size <= replay_buffer._replay_capacity
), f"It's not true that 0 < {desired_size} <= {replay_buffer._replay_capacity}."
assert replay_buffer.size < desired_size, (
f"Replay buffer already has {replay_buffer.size} elements. "
f"(more than desired_size = {desired_size})")
logger.info(
f" Starting to fill replay buffer using random policy to size: {desired_size}."
)
random_policy = make_random_policy_for_env(env)
post_step = add_replay_buffer_post_step(replay_buffer, env=env)
agent = Agent.create_for_env(env,
policy=random_policy,
post_transition_callback=post_step)
max_episode_steps = env.max_steps
with tqdm(
total=desired_size - replay_buffer.size,
desc=
f"Filling replay buffer from {replay_buffer.size} to size {desired_size} using random policy",
) as pbar:
mdp_id = 0
while replay_buffer.size < desired_size:
last_size = replay_buffer.size
max_steps = desired_size - replay_buffer.size - 1
if max_episode_steps is not None:
max_steps = min(max_episode_steps, max_steps)
run_episode(env=env,
agent=agent,
mdp_id=mdp_id,
max_steps=max_steps)
size_delta = replay_buffer.size - last_size
# The assertion below is commented out because it can't
# support input samples which has seq_len>1. This should be
# treated as a bug, and need to be fixed in the future.
# assert (
# size_delta >= 0
# ), f"size delta is {size_delta} which should be non-negative."
pbar.update(n=size_delta)
mdp_id += 1
if size_delta <= 0:
# replay buffer size isn't increasing... so stop early
break
if replay_buffer.size >= desired_size:
logger.info(
f"Successfully filled replay buffer to size: {replay_buffer.size}!"
)
else:
logger.info(
f"Stopped early and filled replay buffer to size: {replay_buffer.size}."
)
def fill_replay_buffer(env: Env, replay_buffer: ReplayBuffer,
desired_size: int):
""" Fill replay buffer with random transitions until size reaches desired_size. """
assert (
0 < desired_size and desired_size <= replay_buffer._replay_capacity
), f"It's not true that 0 < {desired_size} <= {replay_buffer._replay_capacity}."
assert replay_buffer.size < desired_size, (
f"Replay buffer already has {replay_buffer.size} elements. "
f"(more than desired_size = {desired_size})")
logger.info(f"Starting to fill replay buffer to size: {desired_size}.")
random_policy = make_random_policy_for_env(env)
post_step = add_replay_buffer_post_step(replay_buffer, env=env)
agent = Agent.create_for_env(env,
policy=random_policy,
post_transition_callback=post_step)
max_episode_steps = get_max_steps(env)
with tqdm(
total=desired_size - replay_buffer.size,
desc=
f"Filling replay buffer from {replay_buffer.size} to size {desired_size}",
) as pbar:
mdp_id = 0
while replay_buffer.size < desired_size:
last_size = replay_buffer.size
max_steps = desired_size - replay_buffer.size - 1
if max_episode_steps is not None:
max_steps = min(max_episode_steps, max_steps)
run_episode(env=env,
agent=agent,
mdp_id=mdp_id,
max_steps=max_steps)
size_delta = replay_buffer.size - last_size
assert (
size_delta >=
0), f"size delta is {size_delta} which should be non-negative."
pbar.update(n=size_delta)
mdp_id += 1
if size_delta == 0:
# replay buffer size isn't increasing... so stop early
break
if replay_buffer.size >= desired_size:
logger.info(
f"Successfully filled replay buffer to size: {replay_buffer.size}!"
)
else:
logger.info(
f"Stopped early and filled replay buffer to size: {replay_buffer.size}."
)
def __iter__(self):
self.env.reset()
for i in range(self.num_episodes):
trajectory = run_episode(
self.env, self.agent, max_steps=self.max_steps, mdp_id=i
)
yield trajectory.to_dict()
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 train_policy(
env: EnvWrapper,
training_policy: Policy,
num_train_episodes: int,
post_step: Optional[PostStep] = None,
post_episode: Optional[PostEpisode] = None,
use_gpu: bool = False,
) -> np.ndarray:
device = torch.device("cuda") if use_gpu else torch.device("cpu")
agent = Agent.create_for_env(
env,
policy=training_policy,
post_transition_callback=post_step,
post_episode_callback=post_episode,
device=device,
)
running_reward = 0
writer = SummaryWriter()
with summary_writer_context(writer):
train_rewards = []
with trange(num_train_episodes, unit=" epoch") as t:
for i in t:
# Note: run_episode also performs a training step for the agent, if specified in post_step
trajectory = run_episode(env=env, agent=agent, mdp_id=i, max_steps=200)
ep_reward = trajectory.calculate_cumulative_reward()
train_rewards.append(ep_reward)
running_reward *= REWARD_DECAY
running_reward += (1 - REWARD_DECAY) * ep_reward
t.set_postfix(reward=running_reward)
logger.info("============Train rewards=============")
logger.info(train_rewards)
logger.info(f"average: {np.mean(train_rewards)};\tmax: {np.max(train_rewards)}")
return np.array(train_rewards)
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 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 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 = 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 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)
